Method and system for providing a decorative and / or tactile pattern on a decorative panel

Abstract

The invention relates to a method for providing a decorative and / or tactile pattern on a decorative panel, comprising the following steps, a) providing at least one image vector embedding, b) generating at least one embodiment of a desired decorative pattern from random noise guided by said at least one image vector embedding, c) upsampling the at least one embodiment of the desired decorative pattern guided by said at least one image vector embedding at least once to obtain at least one high-resolution decorative pattern, d) separating the high-resolution decorative pattern into at least one colour separation file and / or at least one tactile map, and e) printing at least part of the at least one colour separation file and / or the at least one tactile map on at least a part of an upper surface of a sheet, such that a printed sheet is obtained comprising at least part of the high-resolution decorative pattern and / or a tactile pattern.

Description

[0001] Method and system for providing a decorative and / or tactile pattern on a decorative panel

[0002] The invention relates to a method and system for providing a decorative and / or tactile pattern on a decorative panel. The invention also relates to a decorative panel comprising a decorative and / or tactile pattern and a covering comprising a plurality of such panels.

[0003] Designs for industrial printing, including those intended for use in decorative panels, such as those for use in decorative flooring, decorative walls, furniture and shelving, are limited due to two main reasons: their natural source material, and the printing method's technical limitations.

[0004] As such designs are traditionally based on scans of natural materials, such as wood and stone, the design size is directly correlated with the amount of visually pleasing source material found, scanned, and processed by an expert designer, and / or the manual production of a plurality of samples processed by distinct processes, in order to identify the most pleasing combination of shades, colours, finishes, coatings, textures and the like. This designer goes through a laborious process to select and enhance specifically those features of the source material that are visually pleasing, with the reject rate of said natural source material often ending up being extremely high. This rejection rate translates into an excessive cost in natural resources, often rare or even protected species of wood, and time and labour. This entire process is not only costly and time-consuming, but also leads every year to the felling of a multitude of often endangered trees and species, and a global industry of materials and often illegal trade.

[0005] For this reason, designs for industrial use are mostly limited to a size of 1 m by 1 ,2m, to max 1 ,3m by 2.9m, with a material cost of one to multiple pallets of source material. Secondly, as the industrial printing process is limited to technical size limitations of the most common rotogravure production method, wherein subsequent round printing cylinders consecutively apply multiple layers of coloured ink onto a flexible substrate to form a printed film, larger designs have historically not been desired or technically possible. This has as a result that many designs for decorative panels are limited in size. When applied on a larger surface area, this causes a visual repeat of certain design features, such as knots or cracks in the case of wood designs, or veins or stains in the case of stone designs, across the installed decorative surface. The smaller the design or the larger the panels, the more acute and obvious this visual repeat becomes. For example, a 1 .2 m2of design yields only 5 unique visuals on a floor panel and 3.8m2yields 12 unique planks on a short floor panel, but only a single unique visual on commercial wall panels. This can be and is usually remedied by using two or more distinctive but mutually compatible designs, but at a much higher cost to the manufacturer, as this involves an increased cost in design and an investment in printing cylinders.

[0006] The technical limitations of the rotogravure process over the past years have been completely overturned by the advent of the digital printing process for use the industrial finishing of decorative panels. This digital printing process allows the inline single-step application of at least one design pattern on the surface of decorative panels, obviating the need for films and / or printing cylinders. This means in practice that there is no more technical limitation on the size of a design, and that the single most important limitation now for pattern repeat is the cost of producing a larger design pattern, both in materials and in human labour.

[0007] It is possible to utilize algorithms to generate realistic-looking wood textures, but these algorithms are unable to consider the full complexity of realistic designs, such as wood species, wood origin, texture, figure, grain, grading, finishing, colouring and the like, nor the semantic or often artistic side of decorative designs based on natural materials finished by an expert craftsperson, let alone produce a creative combination of design characteristics, or a wholly new design based on a description rather than a mathematical expression. It is further possible to train deep machine learning networks such as convolutional neural networks (CNN) to identify or recognize wood species by means of analysis of their microscopic buildup, requiring large datasets of microscopic images combined with accompanying descriptions and labels. However, such CNNs are often confused by irregularities in the provided test samples, are incompatible with processed, stained, coated or finished surfaces, are at this time purely non-generative, meaning they may classify wood species on basis of images of their microscopic buildup, but are unable to generate realistic macroscopic design features. They are further computationally inefficient and slow, due to the sequential, non-parallel nature of CNNs. There is therefore a need for a robust wood and stone species recognition and image generation system that overcomes the problems and limitations of the state of the art.

[0008] There is therefore the need for a way to economically and efficiently produce, design or generate new decorative and / or tactile patterns suitable for digital printing on at least one surface of decorative sheets or panels that are not limited in volume of available source material.

[0009] The object of this invention is to provide a method and / or system which can at least partially fulfil the abovementioned need.

[0010] The invention provides thereto a method for providing a decorative and / or tactile pattern on a decorative panel, comprising the following steps: a) providing at least one image vector embedding, b) generating at least one embodiment of a desired decorative pattern from random noise guided by said at least one image vector embedding, c) upsampling the at least one embodiment of the desired decorative pattern guided by said at least one image vector embedding at least once to obtain at least one high-resolution decorative pattern, d) separating the high-resolution decorative pattern into at least one colour separation file and / or at least one tactile map, e) printing at least part of the at least one colour separation file and / or the at least one tactile map on at least a part of an upper surface of a sheet, such that a printed sheet is obtained comprising at least part of the high-resolution decorative pattern and / or the tactile pattern, and f) optionally separating the printed sheet into multiple decorative panels.

[0011] The method according to the invention enables of the provision of at least one unique, continued high-resolution decorative pattern which can be continuously printed upon a sheet, resulting in that individually distinct decorative panels can be obtained. The desired decorative pattern can for example be a natural pattern, such as a wood design, and an embodiment thereof is generated from random noise guided by at least one provided image vector embedding. The obtained embodiment of the desired decorative pattern is subsequently upsampled, or expanded, by making use of the at least one image vector embedding such that at least one high-resolution decorative pattern is obtained. This at least one high- resolution decorative pattern is then separated into at least one colour separation file and / or at least one tactile map which in particular defines the decorative pattern. The colour separation step can for example include the creation of a plurality of colours and / or a greyscale. The at least one colour separation file and / or the at least one tactile map can subsequently be printed on at least a part of an upper surface of a sheet. A printed sheet is then obtained comprising at least part of the high-resolution decorative pattern and / or the tactile pattern, and in particular a continuous or continued printed sheet is obtained. Optionally, the printed sheet can be separated into multiple decorative panels. Each decorative panel will have a distinctive decorative pattern. This is desirable as the distinctiveness of the panels enhances the aesthetic appeal of the final covering or area made of multiple decorative panels. The method enables the provision of an at least partially artificially generated pattern can be at least one comparatively large digital artificially generated design of at least 10m2, preferably at least 25m2 in size after printing; and / or a plurality of at least partially artificially generated patterns that are generated according to a compact representation that captures the semantic essence of and / or a master input image, which may provide an unlimited number of unique printed designs for the decorative pattern which match the same pattern features and may be installed together for a visually pleasing effect. The method further enables the provision of a product in particular a panel, such as but not limited to a floor panel, a building panel, a wall panel and / or a ceiling panel, or a core layer thereof. The panel can for example be a decorative panel.

[0012] The method according to the invention could for example enable the provision of a printed decorative pattern which is based on a semantic and / or compact representation of an input sequence describing a desired pattern and / or an input of a target design. It is for example possible that the desired decorative pattern is provided in a written and / or spoken format and / or via at least one sample image. The desired decorative pattern can for example be provided in a text or speech describing the type of decorative pattern which is to be provided. In a possible embodiment, the method enables the provision of a print which is artificially generated by prompting a text-to-image generative architecture trained on a data set of textual and / or speech descriptions and corresponding decorative designs. In another possible aspect of this invention, after obtaining the at least one high- resolution decorative pattern using the at least one image vector embedding, the said high-resolution decorative pattern is communicated to a virtual environment processor which is in particular used for augmented reality (AR), virtual reality (VR), and / or mixed reality (MR). The virtual environment processor could display the high-resolution (hi-res) decorative pattern to a user such as a client or a designer via a display device, projector, VR headset, VR goggles, head-worn apparatus with integrated display, or any similar device. This then allows the user to have an immersive experience where the user can view the virtual environment having the floor, wall, or ceiling with the hi-res decorative pattern. The virtual environment can also be configured by the user to generate a personalized floor plan based on the intended use space such as a home or office, and the like. The user can for example tweak the design, colour, gloss level, matte level, texture, or any other visual property of the hi-res decorative pattern to fit the specific environment, by a two-way interaction with the generative model. This would conceivably allow for efficiency in producing an exact number of needed panels and designing them to fit designated space, reducing construction waste (generally 10-20% of floor panels delivered to a job site end up as waste due to design uncertainties at purchase). A display device such as a projector or display showing the designed visual may also be used. The user may also be a designer. In one embodiment, the immersive experience is provided with hardware at a point-of-sale, such as a shop, provided through an internet or immersive experience platform, and / or provided to the consumer to be used at the intended use space. When the user is satisfied with the hi-res decorative pattern and / or hi-res decorative surfaces generated by the generative architecture, said patterns or surfaces can then be translated into a vectorized embedding and printed in the exact colour, size, and set-up suitable for the specific use case by a direct printing module. This level of customization and leanness allows for high levels of efficiency, reduction of excess and waste in installation, reduction of stock levels, and consequently of pollution and CO2 emissions.

[0013] In another related embodiment, the obtained at least one high-resolution decorative pattern, surface, or design according to any embodiment of this invention can be interfaced and / or communicated to at least one industrial internet of things (loT), Al-enabled industrial internet of things (AloT) and / or Industry 4.0 (4IR) production environment. Said system then operates as an intermediary platform wherein the design features of the generated high-resolution decorative pattern, the generated designs, and / or the generated to at least one panel production, processing line, and / or monitoring device are communicated. Said panel production, processing line, and / or monitoring device may comprise at least one manufacturing device or tool used in manufacturing decorative building panels. These include but are not limited to, among others, a coating line, a painted bevel line, ripsaw, double end tenoner or profiling machine, visual defect detection machine, extrusion line, cutting equipment, laminating machine, digital printing line, embossing machine, digital embossing machine, or any machine used in producing the said decorative building panels, aid connection to the loT platform enables a more efficient processing of the panels via the integration of the said devices which is not observed in the traditional panel production line. It is conceivable that design parameters such as but not limited to thickness, size, bevels of the panels and the like, are forwarded to at least one processing step for automated and / or Al-controlled adjustment of equipment. It is further conceivable that the at least one image vector embedding is forwarded to at least one Al visual inspection engine which may then measure, calculate and verify the at least one vector embedding in inspecting the final decor, and / or inspect for visual and technical defects, including but not limited to bevel paint colour, damages, cracks, blotches, streaks, glossy spots, faulty tactile pattern and the like.

[0014] The sheet as provided in step e) could optionally also be divided or separated into multiple subsheets. It is also imaginable that the sheet as provided in step e) is adhered to a core layer. Optionally the combination of the core layer and the sheet could be separated into multiple decorative panels. In a preferred embodiment, the sheet as provided in step e) is formed by a core layer or said sheet forms a core layer of a panel.

[0015] In a possible embodiment, the at least one colour separation file and / or the at least one tactile map is printed directly on at least a part of an upper surface of a sheet, such that a printed sheet is obtained comprising at least part of the high-resolution decorative pattern and / or the tactile pattern. At least part of said printed sheet may form a core layer of a panel. It is also possible that the sheet as provided in step e) is pretreated and / or primed prior to at least one colour separation file and / or at least one tactile map being printed thereon. Further, it is also imaginable that the at least one colour separation file and / or the at least one tactile map is printed on at least a part of an upper surface of a sheet and that said sheet is adhered to a core layer or to a panel. It is imaginable that the at least one colour separation file is printed on at least a part of an upper surface of a sheet to form a printed sheet, after which the printed sheet is at least partially provided with at least one cured and / or at least one curable resin to form an uncured coated sheet and / or that that the at least one tactile map is provided in a greyscale, masking, coating-reactive and / or coatingrepellent substance prior to a curing process, subjected to a curing process, and finally removed. Hence, it is for example possible that the printed sheet is adhered to an upper surface of a panel after step e), and preferably that in step f) the printed sheet and the panel are separated into multiple decorative panels. The printed sheet can be a decorative layer.

[0016] The method according to the present invention enables that the at least one surface of a sheet is continuously printed. In a possible embodiment, steps b) - d) are repeated n times and / or steps e) - f) are repeated m times. Possibly, n and m are both integers and n and m may both be at least 1 . In a further preferred embodiment, m is larger than n (m > n) in particular such that the at least one surface of a sheet is continuously printed.

[0017] The method according to the present invention enables that a continuous extruded sheet can be continuously printed with a unique, non-repetitive decorative pattern. It is possible that the method includes at least one guided iterative refinement process which optimizes the decorative pattern and / or which transforms noise into a meaningful representation. When continuously printed, at least one generative network may extend at least part of the design over the edges of the original design, maintaining a seamless transition ad infinito. In a possible embodiment, this is achieved by means of a diffusion model. Such diffusion model may iteratively spread information across the border of the image, simulating the gradual diffusion of pixels from known regions to unknown regions, guided by at least one image vector embedding. It is further conceivable that a transformer architecture is utilized. A transformer architecture may capture semantic and visual relationships between components present in the visual design. Said transformer architecture may be utilized in tandem with said diffusion model. Utilizing both a transformer architecture and a diffusion model is particularly suited for the visual generation step, as it is particularly computationally efficient, can be trained with limited data, and is able to understand and correctly visualize relationships between complex visual components, such as grain, cracks and knots.

[0018] In another embodiment, the method achieves an optimized decorative pattern through the application of a text-to-image model that is trained on a dataset of rich captions. Said rich captions may comprise synthetically generated captions. Said rich captions are configured to be descriptive and exhibit a significant level of detail. In an example, the rich captions, preferably synthetically generated captions may take the form of elaborate descriptions of an envisioned design, such as a wall or flooring panel installed in a room. These descriptions may encompass various physical features, colors, textures, but also meta-features such as use cases, feelings, locations, design or fashion trends, envoked emotions etc, providing a comprehensive representation of the perceived design. In another embodiment, it is conceived that the model is trained on a hybrid dataset, combining synthetically generated captions with ground truth captions. Ground truth captions comprise descriptions taken from human-written text and is conceived to contribute to a diverse training set. It is conceivable that the text-to-image model is configured to utilize synthetic captions at varying blending ratios, with a preference for high percentages. Preferably, the model utilizes at least 60%, more preferably at least 75%, and even more preferably at least 80% of synthetically generated captions in the training dataset. Furthermore, the model limits the use of ground truth captions to at most 35%, preferably at most 30%, and most preferably at most 20% of the dataset. The inventors have found that the quality, accuracy and richness of the generated designs is substantially improved by at least 80% when utilizing synthetically generated captions vs the state of the art.

[0019] In yet another embodiment, the method comprises a step of modifying at least a portion of the image / design. This may be achieved through at least one inpainting step. Inpainting allows alteration of the design / image without leaving visible traces of the editing process thereby ensuring a seamlessly blended or filled in area in the design / image. The said modification step may be performed during the designing step, before printing the image or after printing and obtaining the printed sheet in step e), where the user or operator manually inspects the design or print quality of the said sheet and instructs the processor to perform the inpainting step, wherein at least one vector can be changed and / or make iterative adjustments in the line. This inpainting process is preferably achieved through a diffusion process, more preferably a reverse diffusion process. Preferably, part of the design that is identified as not fitting the aesthetics or general visual effect by the user or operator is removed from the design, through speech-to-text or a visual Ul, to form a partial design. The user or operator may then specify how to adjust the embedding through textual or speech prompt to form a secondary latent vectorization. The diffusion model may then be prompted to run a reverse diffusion process to iteratively update the noise distribution to recover the missing information, in particular the diffusion model may sample from the unconditional prior distribution which provides a generative baseline, condition the reverse diffusion process on observed parts of the image, and / or guide the inpainting process through said secondary latent vectorization. It may be observed that the original diffusion model remains unchanged, and model architecture or weights are not altered, as it merely modifies the reverse diffusion iterations to incorporate the secondary latent vectorization through a shared embedding space. It is also possible that during the inpainting step, lost or missing parts of the desired decorative pattern can be reconstructed. This also serves as a method to fill missing or damaged visual data in the desired decorative pattern. In another possible embodiment, inpainting utilizes the restoration and enhancement of images to augment the decorative pattern’s overall appeal and quality. In some possible embodiments, the inpainting step performs at least a pattern recognition, contextual analysis, and / or adaptive restoration to effectively enhance the quality of the generated decorative pattern, as well as any generated image, pattern, or design.

[0020] The method according to the present invention enables for the provision of large- scale designs of natural patterns which have a non-repetitive character. It is for example possible that decorative pattern comprises a natural pattern based on a wood and / or stone material. Hence, it is possible that the desired decorative pattern involves a natural pattern. The desired decorative pattern may include a wood and / or a stone pattern. It is for example also imaginable that the decorative pattern involves a herringbone pattern. The printed sheet could for example also be designed such that it visually appears that that the sheets consist of multiple wood- based panels which are installed in a herringbone pattern. The combination of the at least one colour separation file and the at least one tactile map results in that a printed sheet can be obtained comprising a high-resolution decorative pattern and a tactile pattern. Combining a high-resolution decorative pattern and a tactile pattern can result in the visualization of a realistic impression. Applying such realistic pattern on a large scale, for example on a large sheet material, could for example be beneficial for flooring purposes. The tactile pattern may follow at least part of the decorative pattern. Within the context of the present invention, the decorative pattern is in particular a two-dimensional (2D) pattern and the tactile pattern is in particular a three-dimensional (3D) pattern, or achieves a three- dimensional effect through further processing.

[0021] In case the method includes step g), it is possible that a total surface area of all distinct decorative patterns presents on at least two panels and preferably of multiple panels obtained in step g) is at least k times a total surface area of a single panel, wherein k is an integer and wherein k is at least 1 . It is also possible that the total surface area of all distinct decorative patterns presents on a plurality of panels obtained in step g), if applied, is at least 3 m2, preferably at least 6 m2, more preferably at least 10 m2, most preferably at least 20 m2. In this way, the chance of having repeated patterns and thus indistinctive (parts of) panels is significantly reduced. The monotony of the decorative pattern can therefore be reduced resulting in a more dynamic and visually stimulating environment. Additionally, non- repetitive patterns can help hide imperfections or irregularities in the flooring surface, as the eye is drawn to the variety of patterns rather than focusing on specific flaws.

[0022] Step b) and / or step e) may further comprise reviewing and / or adjusting the high- resolution decorative pattern by means of a language model and / or a visual user interface. Hence, it is imaginable that further input is provided based on the printed decorative pattern. It is for example imaginable that the colour, structure, scale and / or further characteristics of the decorative pattern are adjusted. It is possible that the desired decorative pattern as provided in step b) is thereto adjusted. It is for example possible that step e) comprises reviewing of the high-resolution decorative pattern and that step b) comprises adjusting of the desired decorative pattern. This will result in the adjusting of the high-resolution decorative pattern. It is possible that the method includes that after step d) a step of outputting a feature map, a vector embedding, and / or semantic description of the high- resolution decorative pattern is performed, preferably via a neural network. The use of a neural network for the provision of a feature map, a vector embedding, and / or semantic description of the high-resolution decorative pattern can further improve the efficiency and / or accuracy of the method. The neural network can recognize and / or classify parts of the decorative pattern accurately which enables that any further outputting steps can be accurately performed. The method may also include training of the neural network, preferably via deep machine learning. Deep machine learning involves for example utilizing existing designs as well as natural and decorative materials to teach a machine or neural network to recognize the different technical, visual, tactile, edge and surface features of its wood and stone patterns on basis of a text sequence, including but not limited to a description, annotation and / or a label. It is for example possible that the at least one neural network provides the desired output based on textual descriptions. The neural network can be applied for generating and / or optimizing the decorative pattern. The method may further include the fine-tuning, training and / or customizing of the neural network with a domain-specific data.

[0023] It is possible that the printing in step f) is chosen from the group consisting of digital printing, inkjet printing, digital coating embossing, digital lacquer embossing, masking printing, and any combination thereof. These systems benefit of the ability to printed decorative patterns with a high accuracy in a relatively fast and effective manner.

[0024] It is further possible that in step b) generating at least one embodiment of a desired pattern is performed via a diffusion model, a generative adversarial network (GAN), a variational auto-encoder, a combination of a CNN and a DPM (diffusion probabilistic model), a latent diffusion model, an LLM-based diffusion model, a U- Net architecture, or any combination thereof. It is also possible that step b) is performed using at least one encoder, at least one decoder, and / or a combination of at least one encoder and at least one decoder, more preferably comprises at least one transformer, most preferably comprises a bimodal transformer model with a text / image shared embedding space. In a possible embodiment, the applied generative adversarial network (GAN) is a combination or at least one GAN variant such as, but is not limited to, conditional GANs (cGAN), adversarial autoencoder, dual GAN (DGAN), stack GAN (StackGAN), cycle GAN (CycleGAN), superresolution GAN (SRGAN), deep convolutional GAN (DOGAN), Wasserstein GAN (WGAN), energy-based GAN (EBGAN), and mode regularized GAN (MRGAN). The CGAN is a supervised learning technique using both labelled and unlabelled data to train the GAN which then improves the accuracy of image generation and text-to-image synthesis. For DGAN, typically two networks are trained in simultaneously with two sets of unlabelled images as input, one network for generating images and the other for discriminating between generated images and real images. The stack GAN produces more realistic or high-resolution images by using multiple generators that are stacked together. CycleGAN translates one image domain to another for an automated image-to-image translation model. In SRGAN, high-resolution images can be generated from low-resolution inputs by applying a deep network in combination with an adversary network to increase the resolution of input data. Deep convolutional neural networks can be used for the generator and the discriminator such as in the DCGAN wherein the GAN comprises convolutiondeconvolution layers. EBGAN uses an energy function to measure the similarity between the real and the generated images and uses the measured similarity to improve the accuracy of image generation. For MRGAN, typically at least one mode regularizer is used to encourage the generator to generate images from all modes of the data distribution. At least one of the above GAN variations or combinations thereof can be used by the processor to generate at least one embodiment of the desired pattern or to perform image or pattern generation that can be printed and used for a decorative panel.

[0025] In another possible embodiment, step b) of generating at least one embodiment of a desired pattern is performed via a latent variable generative model, in particular a diffusion model, most preferably a data-generalization enhancement model, such as but not limited to denoising diffusion probabilistic models (DDPM) and upsampling diffusion probabilistic model (UDPM). The method according to the invention may include that step a) of providing at least one image vector embedding comprises the following substeps: a1) training a text-vector-to-image-vector generative architecture using a plurality of text / image embedding pairs to enable the architecture to convert text embeddings to image embeddings; and / or a2) training a language model to output text embeddings describing natural patterns; and / or a3) providing a textual description of a desired natural pattern to the language model to output a text embedding for the textual description; and / or a4) converting the text embedding to an image vector embedding by using the text-vector-to-image-vector generative architecture.

[0026] Hence, it is possible that the method includes the training of a machine learning network to generate decorative patterns for use in a printing system in particular from noise through a guided iterative refinement process. The teaching and / or training of the machine learning network to recognize and categorize stone and / or wood species of at least one provided sample could also be applied within the method according to the present invention. It is imaginable that the method includes a freezing step which suspends at least a part of the language model and / or the neural network. This may be used to finetune or selectively train certain parts of the language model and / or the neural network while keeping the other parts unchanged to achieve at least a satisfactory accuracy level. In a possible embodiment, said satisfactory accuracy level is preferably at least 95%. In another preferred embodiment, said accuracy level computation is preceded and / or replaced by calculating and / or optimizing a loss function, preferably a loss function chosen from the group of gradient descent, Adam, and stochastic gradient descent (SGD). Said optimization may comprise iteratively adjusting an optimal set of model parameters, weights and biases, that minimize said loss function, in at least one training step, preferably at least 10,000 training steps.

[0027] It is conceivable that the at least one descriptive text for any such existing design comprises any combination of natural and / or artificial characteristics, described in detail and / or in general terms, and / or describe the cultural connotations and / or historical uses of such designs and / or materials. It is for example possible that textual description describes wood-specific physical characteristics. Different species of wood can have significant variations in how wood fibers, cells, and growth rings are arranged, contributing to a distinct grain pattern and texture. Hardwoods further comprise vessels, while softwoods have a more uniform texture due to the absence of vessels. The wood’s species impact further the presence of early- and latewood, sapwood demarcation, luster and gloss, fiber orientation, and how it reacts to surface finishes such as coatings and / or stains. Wood of the same species may further have color differences due to soil infiltrations or extractives that interact with the cellulose of its cell walls and the lignin that bonds them together. Various species react to the infiltrates in diverse ways, while even within a species the wood's color can vary, such as oak harvested in Europe vs Russia, or the US (United States) will have distinct colors. Hence, it is possible that the textual description includes indications of the wood species and / or wood original related visual characteristics. The same applied for the texture and visual characteristics of the wood pattern such as the orientation of the wood cells. It is also possible that the input can be provided for altering the surface texture and appearance of the natural pattern, for example by visualizing treatments like stains or finishes and treatments such as brushing, sawing, sanding, scraping, changing the colour of the wood by means of paint-based processes such as staining, whitewashing and / or washing but also the visualization of physical processes such as smoking or burning, or the application of coatings, waxes or oils. The same teaching could be applied for other natural materials and / or patterns such as stone and / or ceramic patterns.

[0028] It is possible at least one embodiment of a desired decorative pattern generated in step b) is used for training the text-vector-to-image-vector generative architecture in step a1). In this embodiment, steps a2) - a4) are preferably repeated. Steps a2) - a4) can for example be at least two times repeated, preferably multiple times repeated. Repeating the process steps may improve the accuracy and reliability of the method. In a possible embodiment, at least one text-to-image generative architecture, in particular at least one text-vector-to-image-vector generative architecture comprises a multimodal neural network. In a possible embodiment, said multimodal neural network is chosen from the group of Contrastive Language- Image Pretraining (CLIP), Imagen, and / or a Pathways Autoregressive Text-to- Image model (Parti). The type of model can be chosen based upon the further requirements of the method. CLIP uses a relatively complex architecture which vectorizes a plurality of images and text sequences through an embedded transformer model, assigns a similarity index for all vectors, adjusts weights to minimize at least one loss function, and freezes the model. When prompted with a text description, said model can, in a first step, generate an image latent vector representation by means of a prior trained on said weights (conceivably a relatively simple CNN, for the object of the invention more preferably a guided diffusion model), and, in a second step, generate, upsample and / or super-resolve an image by means of a guided diffusion step. Parti uses a relatively complex architecture that tokenizes and vectorizes both texts and images through multi-headed attention transformers. It learns the semantic alignment between the two modalities and then infers an image using a variational auto-encoder. Typically, the image inference is performed through a diffusion model. Imagen model uses a relatively simple architecture which tokenizes and vectorizes text prompts through a multi-headed attention transformer model to create an image from noise through a diffusion model, then super-resolves said image through a guided diffusion process on basis of said vectorized text prompts. This approach can be effective if the text dataset is sufficiently large. It is

[0029] Step a1) of training a text-to-image generative architecture, in particular text-vector- to-image-vector generative architecture may optionally comprise the following steps: ala) providing a dataset comprising a plurality of images and a corresponding plurality of textual descriptions, and al b) generating a plurality of vector embeddings based on the plurality of images and the plurality of textual descriptions.

[0030] It is possible that the dataset comprises a plurality of images and a corresponding plurality of textual descriptions in the form of text / image (embedding) pairs. The text descriptions of the images can contribute to further specifying the output. The text / image (embedding) pairs may for example describe the relationships between a compact representation of at least one text description and a compact representation of at least one image. It is also possible that step al b) of generating a plurality of vector embeddings comprises a step of adjusting parameters and / or weights in at least one text encoder and / or design encoder to contrastively optimize a (cosine) similarity between the plurality of textual descriptions and the plurality of images through backpropagation and / or a loss function convergence.

[0031] In a possible embodiment, the text-vector-to-image-vector generative architecture and / or the language model comprises at least one transformer architecture comprising at least one self-attention layer, preferably a plurality of self-attention layers. The transformer architecture is particularly interesting as it can rely on a self-attention mechanism, which allows the model to weigh the importance of different components of the input sequence, capturing long-range dependencies more effectively than traditional recurrent, adversarial or convolutional neural networks, such as GAN. Using transformers is particularly advantageous over GANs as they allow control over generated content by allowing textual prompts and conditional image synthesis, excel at long-range dependencies, and are able to “understand” semantics and align textual descriptions with visual content. GANs may conceivably generate images based on latent vectors, but controlling specific aspects, such as style, content, and exact location of components, are challenging. Moreover, using language-based instructions with GANs is difficult as these lack the self-attention mechanisms inherent in transformers. Furthermore, it is conceivable that a plurality of transformers are in parallel configuration to form deeper and more powerful models and / or improve computational efficiency while maintaining contextual awareness. By having this parallel configuration, the workload such as those relating to Al processes is divided into a plurality of transformers thus achieving the benefits of parallel computation. As a result, the model is configured to capture more complex patterns and dependencies in the data, while improving computational efficiencies.

[0032] It is possible that after step al a) and prior to step al b) the plurality of images is patchified and / or feature extracted and wherein the plurality of textual descriptions is preprocessed and / or tokenized. It is possible that this step is performed by an encoder, for example a design encoder. It is conceivable that the plurality of images is preprocessed and / or patchified before the at least one feature extraction step. This patchification process allows the model to understand the content and context of the design and align it with text descriptions in a shared embedding space, if present. It is conceivable that at least one neural network is used for at least one feature extraction step. It is also conceivable a feature extraction step is achieved using at least one convolutional neural network (CNN) and / or vision transformer (ViT). The CNN or ViT can be designed to process images and consists of multiple layers of filters that scan the image and identify patterns such as edges, corners, and shapes. The output of each layer is then fed into the next layer, which extracts more complex features. These patterns may then be combined to form higher-level features such as textures and shapes. If patchified, the design may be divided into a grid of patches corresponding to small regions of the design, with each image patch being passed through the feature extraction CNN and later aggregated to form a single latent vector representation for the entire image. An aggregated compact image representation and / or a compact image representation may then be placed in a shared embedding space with a compact text representation, allowing for direct comparison and retrieval between text and images. In a particularly computationally efficient and visually effective embodiment, a ViT can be used for this step, which may patchify and embed each patch in a vector representation. ViTs may be particularly suitable as these are able to learn rich representations by considering the global context of the image, as well as relationships between components present in the image, by means of self-attention mechanisms present in their architecture.

[0033] It is also possible that a patch embedding and / or an attention-based embedding is utilized to create at least one image embedding. The image may then first be divided into a set of patches, such as but not limited to 14x14 patches, each of which is then embedded into a higher-dimensional vector, such as but not limited to a 256-dimensional vector. Said patch embedding may be achieved using a ResNet50 model, and / or a convolutional neural network that has been pre-trained on an image dataset. These patch embeddings may then be processed by at least one attention mechanism to learn the importance and relationship of each patch. The patches may then be weighted and averaged to create a single 256- dimensional image embedding through a flattening and / or normalizing function. The image and text embeddings may then be compared using a contrastive loss function. It is conceivable that the contrastive loss function is designed to maximize the similarity between the image and text embeddings when they are paired correctly and minimize the similarity between the image and text embeddings when they are paired incorrectly. This training objective helps the model to learn to associate images with their corresponding text descriptions. The decorative designs may include visual scans including visual, color, greyscale information as well as tactile and depth information, such as an embossing, texture, edge bevel, and the like.

[0034] It is conceivable that the plurality of images and / or the plurality of textual descriptions is preprocessed. This may comprise the steps or operations of resizing the images to a consistent size, converting them to grayscale or RGB format, dehazing, contrast adjustment, image enhancement, image restoration, image segmentation, object detection, image compression, image manipulation, color space conversion, and / or normalizing of pixel values. As distinctive designs may have different pixel value ranges, a pixel value normalizing process refers to the process of scaling the pixel values of an image to a common scale. Different pixel value ranges can make it difficult for machine learning models to learn from images because the models may prioritize certain features over others. Normalization may then be done by subtracting the mean pixel value and dividing it by the standard deviation, resulting in an image with a mean of zero and a standard deviation of one. For example, if one image has pixel values that range from 0 to 255, while another image has pixel values that range from 0 to 1 , a model that is trained on both images may prioritize the high-intensity features in the first image and ignore the low-intensity features in the second image. Normalizing pixel values to a common range can help to prevent this by ensuring that all images are treated equally and that the model learns from all features in the design.

[0035] Possibly, the plurality of vector embeddings generated in step al b) is the output of a diffusion, U-net and / or an autoregressive model, wherein the input is a text vector embedding. A diffusion process can for example be applied to super resolve the generated decorative pattern. It is conceivable that the diffusion process comprises a series of diffusion steps guided by the compact representation which serves as input for the decorative pattern generation. It’s conceivable a further generative adversarial network is utilized to reduce a loss function that quantifies the differences between the source prompt or image, being an intended image, and the final generated image. In yet another possible embodiment, the plurality of patchified and / or feature extracted images and / or the plurality of preprocessed and / or tokenized textual descriptions is encoded by means of at least one transformer to generate at least one text / image embedding pair.

[0036] Optionally, after generating the at least one embodiment of a desired decorative pattern in step b) and prior to obtaining the at least one high-resolution decorative pattern in step c), the method comprises the step of: processing the decorative pattern by a virtual environment processor, and preferably integrating the decorative pattern into an augmented reality (AR), virtual reality (VR), and / or mixed reality (MR) environment. This way, the method, via the processing step, enables a user to visualize the high-resolution decorative pattern applied to a simulated surface within a virtual space, and to modify at least one visual property of the high- resolution decorative pattern through a two-way interaction with a generative model.

[0037] In a possible embodiment, step a1) and / or step al b) comprises at least one image processing step at least partially performed by at least one visual transformer (ViT), and at least one text processing step at least partially performed by at least one transformer, preferably in a shared embedding space. Said at least one visual transformer and at least one transformer comprise at least one self-attention layer, preferably a plurality of self-attention layers. This allows for correct semantic relations between text components, image components, and correct training of the image / text contrastive model.

[0038] In a further possible embodiment, step a4), step b) and step c) are at least partially performed by means of a diffusion model, preferably a DDPM and / or UDPM.

[0039] The method according to the invention may also include the step of providing at least one resin, in particular at least one reactive resin, bio-based resin, hotmelt resin, thermo-set resin, UV-set resin, EB-cured resin, excimer-cured resin and / or combinations thereof onto an upper surface of the printed sheet. The at least one resin may form a wear layer, possibly at least partially comprising abrasive particles. The method may further include the step of curing said at least one resin. The invention also relates to a system which is in particular configured for performing the method according to the present invention, said system comprising:

[0040] - at least one processor,

[0041] - at least one printer, preferably at least one digital printer, more preferably at least one industrial digital printer, electronically connected to the at least one processor, and

[0042] - at least one conveyor for conveying a printable sheet through the printer, in particular the digital printer; wherein the processor is preferably selected from artificial intelligence processing unit (AIPU), graphics processing unit (GPU), tensor processing unit (TPU), field programmable gate arrays (FPGA), Al-specific ASIC, neuromorphic processor, quantum processor, or any combination thereof.

[0043] The system according to the invention is in particular configured to perform any of the embodiments according to the present invention. The described embodiments of the method could also be implemented in the system according to the invention. The system is in particular a system for providing a decorative sheet or panel preferably provided with a decorative and / or tactile pattern. The conveyor could for example comprise multiple rollers configured for guiding at least one sheet.

[0044] The system could further comprise at least one separator, which may optionally comprise at least one blade and / or at least one saw, for separating a sheet into multiple panels and optionally at least one tenoner or at least one profiling machine for cutting at least a part of the sheet, preferably at least a part of the side edges of the sheet. It is also possible that the at least one profiling machine is configured to provide at least one panel with at least one bevel, such as a cut bevel, coated bevel and / or painted bevel. The profiling machine may also be configured to provide at least one panel, and in particular at least two opposing side edges thereof, with complementary coupling means. In this way, the system can be applied for the production of decorative panels. It is for example possible that the system is configured for attaching the sheet to at least one core layer, in particular prior to the separating step, if applied.

[0045] The system may comprise at least one virtual environment processor configured to process a high-resolution decorative pattern and to integrate said high-resolution decorative pattern into an augmented reality (AR), virtual reality (VR), and / or mixed reality (MR) environment. By providing such virtual environment processor, the system, in particular via said virtual environment processor, enables a user to visualize and / or modify at least one visual property of the high-resolution decorative pattern. Preferably, said virtual environment processor allows for a two-way interaction with a generative model and transmit the finalized modified high- resolution decorative pattern to a direct printing module.

[0046] The system may also comprise at least one extruder. The extruder is preferably configured for continuous extrusion of a sheet, in particular a printable sheet. It is for example possible that the extruder is configured for the production of a sheet having a thickness between 0.01 and 30 mm.

[0047] The processor is preferably selected from the group of: artificial intelligence processing unit (AIPU), graphics processing unit (GPU), tensor processing unit (TPU), field programmable gate arrays (FPGA), Al-specific ASIC, neuromorphic processor, quantum processor, or any combination thereof. In a possible embodiment, the processors can be a plurality of processors that are stacked or configured to operate cooperatively to perform parallel computing. For example, a plurality of GPUs can be stacked to do multiple computations simultaneously. It is also conceivable multiple instances of the model may run in parallel on a plurality of processors and / or edge devices comprising distinct processors. The processor is in particular an internal processor. The processor as applied in the system could optionally be an external processor. These processors are in particular suitable for applying a method according to the present invention. It is also conceivable that the processor is an edge device or is comprised in an edge device. In this way, the edge device can provide the required with the further system.

[0048] In a possible embodiment, the processor comprises at least one Artificial Intelligence Processing Unit (AIPU) or at least one device configured for efficient processing of artificial intelligence (Al) workloads or processes. Herein, the AIPU processes specialized instructions and operations designed for common Al tasks. Since, Al processes require complex mathematical computations that can be processed simultaneously, it is preferred that the AIPU can handle parallel processing wherein multiple processes or AlPUs are used to improve the overall efficiency of the invention. The AIPU is also preferred to be designed for handling large amounts of data or has high throughput since the training of Al generally requires large amount of training data.

[0049] In another possible embodiment, the processing unit comprises a plurality of graphics processing units (GPU) and has a performance of at least 5, preferably at least 10, most preferably at least 30 petaflops. The AIPU preferably has at least 640 GB of GPU memory and / or , at least 1TB, most preferably at least 2 TB of system memory for increased training and / or generation performance. The GPUs are used as part of the processing unitor can also be a stand-alone or independent processing unitAnd are specifically suitable for the envisioned method.

[0050] In yet another possible embodiment, the processor comprises at least one tensor, tensor core or tensor processing unit (TPU) for accelerating the training step involved in Al-related processes. Moreover, TPUs also speed up the execution of complex machine learning (ML) models which are fundamental to several Al applications. These results are achieved due to TPU’s specialized design allowing rapid and efficient matrix vector operations .

[0051] In yet a further possible embodiment, the processor is configured to perform text data processing, speech-to-text processing, and / or natural language processing (NLP) for performing the text-vector-to-image-vector generative architecture or text- to-image processes as well as the operations, computations, conversions, or issuing instructions relating to text data such as the textual descriptions used in generating or are related to the plurality of images and / or models. In this scenario, the processor performs text data processing to make the text data more suitable for analysis and / or modelling. Moreover, the text data processing can also be performed to remove or reduce noise, make uniform the structure of the text data through text segmentation, normalization, case conversion, stemming, lemmatization, stopword or punctuation removal, spelling correction, and / or combinations thereof.

[0052] The system may comprise a first application programming interface connected to the processor, wherein the first application programming interface is optionally connected to a large language model, a database comprising vector embeddings, a model, and / or an image storage. The system may also comprise a second application programming interface connected to the first application programming interface, wherein the second application programming interface is connected to a user interface, a database comprising a plurality of images, a database comprising a plurality of textual descriptions corresponding to the plurality of images, and / or a model.

[0053] At least one printer, preferably at least one digital printer, more preferably at least one industrial digital printer, is in particular configured for printing at least part of at least one colour separation file and / or at least one tactile map on at least a part of an upper surface of a sheet, such that a printed sheet is obtained comprising at least part of the high-resolution decorative pattern and / or a tactile pattern. As indicated above, preferably after the sheet has been provided with a decorative and / or tactile pattern, the sheet can be cut into multiple panels. It is imaginable that at least one panel, and preferably each panel is provided with a bevel. The system may be configured for printing at least part of at least one colour separation file and / or at least one tactile map on at least a part of at least one bevel of at least one panel.

[0054] The system according to the invention may also be configured to provide at least one resin, in particular at least one reactive resin, bio-based resin, hotmelt resin, thermo-set resin, UV-set resin, EB-cured resin, excimer-cured resin and / or combinations thereof onto an upper surface of the printed sheet. The at least one resin may form a wear layer, possibly at least partially comprising abrasive particles. The system may also be configured to provide a masking, reactive, and / or resin-repelling substance according to the tactile feature map and / or greyscale determined by the system and / method according to the invention. The system may additionally comprise at least one curing unit for curing, reacting, setting and / or activating said at least one resin and / or at least one substance. Optionally, the system, and in particular at least one structuring unit, may be configured for removing at least part of the cured, reacted, set and / or activated resin and / or substance to form an at least partially textured and / or embossed surface. In another possible embodiment, the system and / or method according to the invention may include at least partially compressing and / or imprinting said at least one resin in particular by means of mechanical imprinting device to form an at least partially textured or embossed surface.

[0055] The system and method according to the present invention provide for more control of the production of sheets and / or panels. It is for example conceivable that the method and / or system are configured for producing a batch of sheets or panels based on optimized sizes and installation formats. This may reduce waste and excess, for example for flooring. It is conceivable that specific floors are pre-cut and in correct size on the basis of the predefined floor area which will speed up installation tremendously and reduce waste. Hence, it is conceivable that not all panels as obtained by applying the method and / or system according to the present invention have similar dimensions, but that the dimensions are predetermined based on the final covering which is to be made from said panels. A panel recipe could be determined and performed by the method and / or system according to the present invention such that a specific batch of panel for a specific purpose is obtained.

[0056] The invention also relates to a decorative panel, comprising core layer comprising at least one printed decorative pattern, in particular a digitally printed decorative pattern, wherein the printed decorative pattern is digitally and / or artificially generated. The printed decorative pattern preferably resembles a wood or stone pattern. The panel can for example be a floor panel, a building panel, a wall panel and / or a ceiling panel, or a core layer thereof. The panel can for example be a decorative panel.

[0057] It is possible that the panel comprises at least part of a sheet as referred to in the method and / or system according to the present invention. It is also possible that the panel is formed by such sheet. The decorative pattern can be printed directly on the core layer or upon a sheet provided upon the core layer. The sheet can be an extruded sheet and / or the core layer can be an extruded core layer. Any suitable material could be applied for the core layer and / or the sheet, for example at least one polymer material or at least one composite material comprising at least one mineral filler and / or at least one polymeric binder. The core layer and / or sheet could be at least partially foamed. In yet another possible embodiment, the invention relates to a decorative sheet comprising core layer comprising at least one printed decorative pattern, in particular a digitally printed decorative pattern, wherein the printed decorative pattern is digitally and / or artificially generated. The printed decorative pattern preferably resembles a wood or stone pattern.

[0058] In a further possible embodiment, the printed decorative pattern comprises at least one and / or a tactile pattern and / or a surface texture. It is imaginable that at least part of the at least one tactile pattern is applied conform at least part of the printed decorative pattern. In a preferred embodiment, the printed decorative pattern is digitally and / or artificially generated via method according to the present invention. The panel may be obtained by using a system according to the present invention.

[0059] The panel according to the present invention may comprise at least one backing layer optionally comprising at least one ply of resin impregnated paper, wherein at least one ply of paper is impregnated with a resin composition comprising at least one thermosetting resin and / or at least one polymer adhesive. The backing layer, if applied, can be directly attached to the bottom core surface of the core layer. The backing layer can for example be a balancing layer.

[0060] It is conceivable that the panel and in particular the core layer comprises at least one reinforcing layer. In a possible embodiment, the core layer comprises multiple core layers wherein optionally two adjacent core layers enclose a reinforcing layer. The presence of at least one reinforcing layer may further enhance the impact resistance of the core layer, and thus the panel. At least one reinforcing layer may for example be present in the form of a reinforcing mat, a membrane and / or a mesh. At least one reinforcing layer may for example comprise fiber glass, polypropylene, jute, cotton and / or polyethylene terephthalate.

[0061] It is possible that the panel is substantially flat and comprises at least two opposing side edges, wherein each of the at least two opposing side edges comprises complementary coupling means, and wherein the complementary coupling means are configured to a plurality of the decorative panels together. Depending on the intended purpose of the panel, the panel, and in particular the core layer thereof may comprise coupling means. In a possible embodiment, the panel comprises two pairs of opposite side edges, wherein at least one pair of opposite side edges, and preferably each pair of opposite side edges, is provided with complementary coupling parts. The core layer of the panel according to the present invention may comprise at least one pair of opposing (side) edges, said pair of opposing (side) edges comprising complementary coupling parts configured for mutual coupling of adjacent panels. The coupling parts may form part of the core layer. The coupling parts of the panel may for example be interlocking coupling parts, which are preferably configured for providing both horizontal and vertical locking. Interlocking coupling parts are coupling parts that require elastic deformation, a click or a movement in multiple directions to couple or decouple the parts with or from each other. Any suitable interlocking coupling parts as known in the art could be applied. A non-limiting example is an embodiment wherein a first edge of said first pair of opposing edges comprises a first coupling part, and wherein a second edge of said first pair of opposing edges comprises a complementary second coupling part, said coupling parts allowing a plurality of panels to be mutually coupled; wherein the first coupling part comprises a sideward tongue extending in a direction substantially parallel to a plane defined by the panel, and wherein the second coupling part comprises a groove configured for accommodating at least a part of the sideward tongue of another panel, said groove being defined by an upper lip and a lower lip. It is conceivable the complementary coupling parts require a downward scissoring motion when engaging, or are locked together by means of a horizontal movement. It is further conceivable that the interconnecting coupling parts comprise a tongue and a groove wherein the tongue is provided on one side edge of one pair of opposing side edges, and the groove is provided on the other side edge, or an adjacent side relative to that of the tongue, of the same pair of opposing side edges.

[0062] The invention further relates to a covering comprising a plurality of decorative panels according to the invention, wherein all digitally printed decorative patterns of each decorative panel are distinct. The covering may for example be a floor covering, a wall covering or a ceiling covering. It is for example possible that the covering comprises a substantially planar top surface and a substantially planar bottom surface, and wherein a surface area of the top surface is at least 4 m2, preferably at least 6 m2, more preferably at least 10 m2, even more preferably at least 20 m2, most preferably at least 50 m2. It is also possible that the covering according to the invention is specially designed for a predetermined purpose. The covering may comprise a plurality of panels which are designed based on optimized sizes and installation formats. This may reduce waste and excess.

[0063] It will be clear that the invention is not limited to the exemplary embodiments which are described here, but that countless variants are possible within the framework of the attached claims, which will be obvious to the person skilled in the art. In this case, it is conceivable for different inventive concepts and / or technical measures of the above-described variant embodiments to be completely or partly combined without departing from the inventive idea described in the attached claims.

[0064] The verb 'comprise' and its conjugations as used in this patent document are understood to mean not only 'comprise', but to also include the expressions

[0065] 'contain', 'substantially contain', 'formed by' and conjugations thereof.

Claims

Claims1 . Method for providing a decorative and / or tactile pattern on a decorative panel, comprising the following steps: a) providing at least one image vector embedding; b) generating at least one embodiment of a desired decorative pattern from random noise guided by said at least one image vector embedding; c) upsampling the at least one embodiment of the desired decorative pattern guided by said at least one image vector embedding at least once to obtain at least one high-resolution decorative pattern; d) separating the high-resolution decorative pattern into at least one colour separation file and / or at least one tactile map; e) printing at least part of the at least one colour separation file and / or the at least one tactile map on at least a part of an upper surface of a sheet, such that a printed sheet is obtained comprising at least part of the high-resolution decorative pattern and / or a tactile pattern; and f) separating the printed sheet into multiple decorative panels.

2. Method according to claim 1 , wherein the printed sheet is adhered to an upper surface of a panel after step e), and wherein in step f) the printed sheet and the panel are separated into multiple decorative panels.

3. Method according to claim 1 or claim 2, wherein steps b) - d) are repeated n times and wherein steps e) - f) are repeated m times, wherein n and m are both integers and wherein n and m are both at least 1 , preferably, wherein m > n.

4. Method according to any one of claims 1- 3, wherein the at least one surface of a sheet is continuously printed.

5. Method according to any one of claims 1 - 4, wherein the decorative pattern comprises a natural pattern based on a wood and / or stone material.

6. Method according to any one of claims 1 - 5, wherein a total surface area of all distinct decorative patterns presents on a plurality of panels obtained in step g)is at least k times a total surface area of a single panel, wherein k is an integer and wherein k is at least 1 , and / or wherein the total surface area of all distinct decorative patterns presents on a plurality of panels obtained in step g) is at least 3 m2, preferably at least 6 m2, more preferably at least 10 m2, most preferably at least 20 m2.

7. Method according to any one of claims 1 - 6, wherein step e) further comprises reviewing and adjusting the high-resolution decorative pattern by means of a language model and / or visual user interface.

8. Method according to any one of claims 1 - 7, comprising after step d) a step of outputting a feature map, a vector embedding, and / or semantic description of the high-resolution decorative pattern via a neural network.

9. Method according to any one of claims 1 - 8, wherein the printing in step f) is chosen from the group consisting of digital printing, inkjet printing, digital coating embossing, digital lacquer embossing, masking printing, and any combination thereof.

10. Method according to any one of claims 1 - 9, wherein step b) of generating at least one embodiment of a desired pattern is performed via a diffusion model, a variational auto-encoder, a combination of a CNN and a DPM, a latent diffusion model, an LLM-based diffusion model, a U-Net architecture, or any combination thereof.11 . Method according to any one of claims 1 - 10, wherein step a) of providing at least one image vector embedding comprises the following substeps: a1 ) training a text-vector-to-image-vector generative architecture using a plurality of text / image embedding pairs to enable the architecture to convert text embeddings to image embeddings; a2) training a language model to output text embeddings describing natural patterns; a3) providing a textual description of a desired natural pattern to the language model to output a text embedding for the textual description;a4) converting the text embedding to an image vector embedding by using the text-vector-to-image-vector generative architecture.

12. Method according to claim 11 , wherein at least one embodiment of a desired decorative pattern generated in step b) is used for training the text-vector- to-image-vector generative architecture in step a1 ), and wherein steps a2) - a4) are repeated.

13. Method according to claim 11 or claim 12, wherein the text-vector-to-image- vector generative architecture is selected from Contrastive Language-Image Pretraining (CLIP), Imagen, or a Pathways Autoregressive Text-to-lmage model (Parti).

14. Method according to any one of claims 11 - 13, wherein step a1) of training a text-vector-to-image-vector generative architecture comprises the following steps: al a) providing a dataset comprising a plurality of images and a corresponding plurality of textual descriptions, and al b) generating a plurality of vector embeddings based on the plurality of images and the plurality of textual descriptions.

15. Method according to claim 14, wherein step al b) of generating a plurality of vector embeddings comprises a step of adjusting parameters and / or weights in at least one text encoder and / or design encoder to contrastively optimize a similarity between the plurality of textual descriptions and the plurality of images through backpropagation and / or a loss function convergence.

16. Method according any one of claims 11 - 15, wherein the text-vector-to- image-vector generative architecture and / or the language model comprises at least one transformer architecture comprising at least one self-attention layer, preferably a plurality of self-attention layers.

17. Method according to any one of claims 10 - 16, wherein after step a1 a) and prior to step a1 b) the plurality of images is patchified and / or feature extracted and wherein the plurality of textual descriptions is preprocessed and / or tokenized.

18. Method according to any one of claims 10 - 17, wherein the plurality of vector embeddings generated in step a1 b) is the output of a diffusion, U-net and / or an autoregressive model, wherein the input is a text vector embedding.

19. Method according to claim 16 or claim 18, wherein the plurality of patchified and / or feature extracted images and / or the plurality of preprocessed and / or tokenized textual descriptions is encoded by means of at least one transformer to generate at least one text / image embedding pair.

20. Method according to any one of claims 1 -19, wherein after generating the at least one embodiment of a desired decorative pattern in step b) and prior to obtaining the at least one high-resolution decorative pattern in step c), the method comprises the step of:- processing the decorative pattern by a virtual environment processor, and;- integrating the decorative pattern into an augmented reality (AR), virtual reality (VR), and / or mixed reality (MR) environment.21 . System for performing the method according to any one of claims 1 - 20, comprising:- at least one processor;- at least one printer, in particular at least one digital printer, electronically connected to the processor; and- at least one conveyor for conveying a printable sheet through the printer, in particular the digital printer; wherein the processor is selected from artificial intelligence processing unit (AIPU), graphics processing unit (GPU), tensor processing unit (TPU), field programmable gate arrays (FPGA), Al-specific ASIC, neuromorphic processor, quantum processor, or any combination thereof.

22. System according to claim 21 , comprising:- at least one separator, optionally comprising a blade or a saw, for separating a sheet into multiple panels; and- at least one tenoner or at least one profiling machine for cutting at least a part of the sheet, preferably at least a part of the side edges of the sheet.

23. System according to any one of claim 21 or 22, further comprising;- at least one virtual environment processor configured to process a high- resolution decorative pattern and to integrate said high-resolution decorative pattern into an augmented reality (AR), virtual reality (VR), and / or mixed reality (MR) environment.

24. System according to any one of claims 21 - 23, comprising at least one extruder configured for continuous extrusion of a printable sheet.

25. System according to any one of claims 21 - 24, wherein the processor is an edge device.

26. System according to any one of claims 21 - 25, comprising a first application programming interface connected to the processor, wherein the first application programming interface is connected to a large language model, a database comprising vector embeddings, a model, and / or an image storage.

27. System according to claim 26, comprising a second application programming interface connected to the first application programming interface, wherein the second application programming interface is connected to a user interface, a database comprising a plurality of images, a database comprising a plurality of textual descriptions corresponding to the plurality of images, and / or a model.

28. Decorative panel comprising a core layer and a printed decorative pattern, wherein the printed decorative pattern is digitally and / or artificially generated, and wherein the printed decorative pattern resembles a wood or stone pattern.

29. Decorative panel according to claim 28, wherein the printed decorative pattern is digitally and / or artificially generated via method according to any of claims 1 - 20.

30. Decorative panel according to claim 28 or claim 29, wherein the printed decorative pattern comprises a tactile pattern.31 . Decorative panel according to any of claims 28 - 30, wherein the panel is substantially flat and comprises at least two opposing side edges, wherein each of the at least two opposing side edges comprises complementary coupling means, and wherein the complementary coupling means are configured to a plurality of the decorative panels together.

32. Covering comprising a plurality of decorative panels according to any one of claims 28 - 31 , wherein all digitally printed decorative patterns of each decorative panel are distinct.

33. Covering according to claim 32, wherein the covering comprises a substantially planar top surface and a substantially planar bottom surface, and wherein a surface area of the top surface is at least 4 m2, preferably at least6 m2, more preferably at least 10 m2, even more preferably at least 20 m2, most preferably at least 50 m2.