Language model assisted generation of images with coherence

Abstract

Systems, apparatuses and methods may provide for technology that conducts a warp operation on an input image to obtain a shifted view and a mask, wherein the mask indicates missing regions in the shifted view associated with the warp operation, conducts, by a first pre-trained artificial intelligence (AI) model, an inpaint operation on the missing regions based on a textual output from a second pre-trained AI model, wherein the textual output is in response to one or more prompts, and repeats the warp operation and the inpaint operation to obtain a plurality of generated views. The technology can also fuse the plurality of generated views into a 360° panoramic view.

Description

[0001] LANGUAGE MODEL ASSISTED GENERATION OF IMAGES WITH COHERENCE

[0002] CROSS-REFERENCE TO RELATED APPLICATIONS

[0003] The present application claims the benefit of priority to U.S. Provisional Patent Application No. 63 / 599,669, filed on November 16, 2023.

[0004] BACKGROUND

[0005] Generating panoramic scenes from a single input image can be challenging. For example, conventional solutions may produce subpar outputs with duplicated objects (e.g., multiple beds in a bedroom) or require time-consuming human text inputs for each view that is used to produce the panoramic scene.

[0006] BRIEF DESCRIPTION OF THE DRAWINGS

[0007] The various advantages of the embodiments will become apparent to one skilled in the art by reading the following specification and appended claims, and by referencing the following drawings, in which:

[0008] FIG. 1 is an illustration of an example of a pipeline according to an embodiment;

[0009] FIG. 2 is an illustration of an example of a pseudo-code listing according to an embodiment;

[0010] FIG. 3 is a comparative illustration of an example of conventional image-to- panorama visualization results and enhanced image-to-panorama visualization results according to an embodiment;

[0011] FIG. 4 is a flowchart of an example of a method of operating a performance- enhanced computing system according to an embodiment;

[0012] FIG. 5 is a block diagram of an example of a performance-enhanced computing system according to an embodiment;

[0013] FIG. 6 is an illustration of an example of a semiconductor package apparatus according to an embodiment;

[0014] FIG. 7 is a block diagram of an example of a processor according to an embodiment; and

[0015] FIG. 8 is a block diagram of an example of a multi -processor based computing system according to an embodiment. DETAILED DESCRIPTION

[0016] The challenge of generating panoramic scenes (“panoramas”) from a single input image is related to the fact that most training data for scene generation is based on pictures with a limited field of view and not fully consistent surrounding data. Existing approaches to extend the field of view may still be insufficient.

[0017] More particularly, some approaches may treat the panorama as a single equirectangular image and learn to generate the image in a single operation through a generation model (e.g., latent diffusion model). This approach struggles to close the loop at both ends of the generated equirectangular image.

[0018] Other approaches create panoramas by generating multiple perspective views using more robust pre-trained diffusion models trained on large-scale perspective data. Under these approaches, multi- view consistency may be achieved by warping or using extra multi-view attention modules fine-tuned on small datasets. These approaches struggle, however, to generate diverse 360° views.

[0019] The technology described herein provides for Language Model Assisted Generation of Images with Coherence (“L-MAG1C”), which leverages large language models (LLMs) to guide the generation of coherent views in 360° panoramic scenes. More particularly, the technology described herein generates a coherent 360° panoramic scene given a single seed image. Embodiments leverage large language models to guide the generation of diverse, yet coherent, views based on the seed image. The term “coherent” herein refers to not only a smooth view transition, but also refers to the capacity to create a scene without unnecessarily repeated objects in multiple views (e.g., multiple beds in a bedroom). The output quality and resolution of panoramic scenes is further enhanced by super-resolution and multi-view fusion techniques.

[0020] Indeed, the technology described herein can generate 360° panoramic views with higher quality and more coherent scene layout. The technology described herein is also effective on “in-the-wild” input images and synthetic data. Additionally, the technology described herein can be combined with mature conditional generative artificial intelligence (Al) models to create panoramic views from diverse input modalities, such as text descriptions of the scene, sketch drawings of a perspective view of the scene, depth maps, and so forth. This feature enables the technology described herein to be applied to applications that involve inputs beyond a single image. Moreover, the realistic and diverse panoramic scenes can be used to create 3D (three- dimensional) point clouds and immersive videos for applications such as video games, virtual reality, and 3D environments.

[0021] When the technology described herein is implemented, redundant objects (e.g., multiple beds in a bedroom) are adaptively eliminated from scenes. Additionally, embodiments use a large language model to generate scene descriptions that are then used as prompts to generate panoramic or environmental scenes. Moreover, the generated scene typically matches with what a large language model would describe.

[0022] As shown in FTG. 1 , an L-MAGIC pipeline 10 generates panoramic scenes using an iterative warp-and-inpaint process. Given a starting image, a warp operation 12 generates a shifted perspective view (e.g., a rotation of viewing angle) and a mask indicating missing regions due to the shift. An inpaint operation 14 fills in the missing regions using image generation models with the assistance of textual outputs from large language models (LLMs, e.g., pre-trained Al models) 16 in response to automated prompts. These text outputs from the LLMs 16 guide the generated regions to be a more diverse and realistic scene. The LLMs 16 are also used to monitor each inpaint operation 14, so that for the rare case where inpainting does not follow the guidance from the LLMs 16, the inpaint operation 14 can be re-run to ensure the effectiveness of the guidance from the LLMs 16.

[0023] The final panorama is created by fusing / merging 20 the generated views into a 360° panoramic view 22 with some post-processing to enhance the quality and resolution. Furthermore, the pipeline 10 uses techniques to distinguish between scene layout regions such as, for example, foreground and background, to identify objects and use the information in the LLM-guided scene generation with positive and negative prompts to enhance the scene layout diversity and consistency expectations of humans. In the illustrated example, the panoramic view 22 is used to generate an immersive video 24 and / or three-dimensional (3D) point cloud 26.

[0024] More particularly, the input to the pipeline 10 is an image / either captured by a camera 28 (e.g., in the real- world) or synthesized by a conditional diffusion model 30 (e.g., ControlNet). Multiple novel views to compose the 360° panoramic view 22 are generated by iterative warping and inpainting. Pre-trained diffusion models assisted by pre-trained language models are used to generate views with both high-quality local textures and coherent 360° layouts. Further quality enhancement techniques ensure smooth blending of multiple views into high-resolution panoramic scenes. The pipeline 10 can also generate the panoramic view 22, the immersive video 24 and the 3D point cloud 26 from various types of inputs, such as textual input 32, synthetic images 34 and / or sketch drawings 36.

[0025] FIG. 2 shows a pseudo-code listing 40 in which, starting from a natural image, a panoramic view / scene is generated via iterative warping and inpainting. The warp operation generates an incomplete perspective view and a mask of the missing region. The inpaint operation completes the masked region with assistance from language models. The final panoramic view is created by fusing the generated views with some post-processing to enhance the quality and resolution.

[0026] More particularly, during the warp operation, L-MAGIC uses image warping to warp existing views to a novel one, where the novel view has a relative rotation compared to existing views and covers an incomplete region of the 360° panoramic scene. Each warp operation takes all existing views and the relative camera rotation between the novel view and the existing views as the input, and produces the image of the novel view and a binary mask indicating whether each pixel of the novel view is complete, where a complete pixel is covered by existing views. No camera translation is used during warping since only perspective views of the panorama are being generated.

[0027] During each warp operation, all completed views are projected onto a unit sphere representing the panoramic scene. Then, the next incomplete view is rendered to the inpaint operation based on the relative camera pose. To project an image to the unit sphere, a mesh is first constructed by defining the vertices V on each image pixel and creating edges E between adjacent pixels. Then, the vertices are projected to a unit sphere by: , ogeneous coordinate of a pixel, and vspis the projected location. To warp a completed view A to a novel view B, where R is the rotation matrix from A to B, each projected vertex vspof A is rotated by vrot = Rvsp, and then rasterization-based rendering is performed (I, M~) = rasterize(Vrot, E, .') where Vrot is the set of rotated vertices and K' is the intrinsic matrix of image B. The output I is a warped image and M is a binary mask indicating whether inpainting is to be performed for a pixel (e.g., obtained by checking for each pixel to determine whether ray-casting intersects a valid mesh face). To ensure the local inpainting consistency of each perspective view, a large field of view (FoV) is used and the rotation angles are adjusted so that both known and unknown regions are reasonably large after warping. In practice, a FoV of 100 degrees with roughly 40° of rotation between adjacent views provides advantageous results. To further reduce the iterative error accumulation, the scene is expanded alternatively from both sides of the input image, rather than expanding the scene in a single direction. To ensure a smooth 360° loop closure, the rotation angles are tuned so that the final view has a relatively large incomplete region at the center, resulting in a sequence of views with rotation angles of {0°, 41°, -41°, 82°, -82°, 123°, 200.5° (e.g., for loop closure)}.

[0028] During inpainting, L-MAGIC uses two-dimensional (2D) image inpainting models such as, for example, a Stable Diffusion v2 inpainting model (e.g., first pretrained Al model) to complete the missing region of the novel view, where the missing region refers to the pixels where the output binary mask of the warp operation have a value of zero.

[0029] More particularly, the inpaint operation completes a warped view with a consistent local style and a coherent 360° scene layout. In one example, the Stable Diffusion v2 inpainting model can effectively extrapolate the large missing region of each warped view while maintaining local style consistency. Naive inpainting, however, without any a priori information (e.g., specific layout-related context information) may generate severe artifacts. For example, one common a priori approach may operate on a user-provided text description of the scene or input image. Using the same description in different views, however, may generate duplicate objects such as multiple beds in a bedroom, since perspective inpainting methods have no mechanism to split the layout into different views. To address these problems, a language model £(■) (e.g., ChatGPT4, second pre-trained Al model) to guide the inpainting process and a vision language model £v(p) (e.g., bootstrapping languageimage pre- training / B LIP-2, third pre- trained Al model).

[0030] The pseudo-code listing 40 demonstrates that before warping and inpainting, £„(■) is first prompted to generate the description di for the input image I (e.g., line 2). Two questions are asked so that di contains both coarse and fine levels of detail. Next, £(■) is asked to imagine the global scene layout dj6o (e.g., line 3) based on di, where each line of d36o corresponds to the description of a specific view. To avoid duplicate objects, a compact description of individual views is requested without mentioning objects in other views.

[0031] The variable d36o primarily contains objects of individual views. Using such descriptions as the inpainting prompt can lead to inconsistent style at distant views. Accordingly, £(■) is asked to remove objects from di and obtain the final scene-level description dscene (e.g., “a bedroom with a wooden bed” becomes “a bedroom” (e.g., line 4)). The variable dscene is later used together with ds6o to ensure a consistent multiview style. Though dscene ensures the multi-view style consistency, the training data bias of diffusion models may still result in objects commonly associated with a particular scene being generated, even if not explicitly mentioned in d 60- For example, a bed is often generated with the word “bedroom” in the prompt, resulting in duplicate beds in multiple views. To solve this problem, £(■) is permitted to automatically determine whether there are some objects in the scene to undergo repetition avoidance (e.g., line 5).

[0032] After each warp operation, the outputs from lines 2 to 5 are used to automatically generate the prompt for text conditioned inpainting (e.g., line 12). Specifically, for the warped view with 0° rotation (i = 1), dscene is used as the prompt (di) for text-conditioned inpainting (e.g., line 7). For other views (e.g., line 13), if there is no object in drepeat (e.g., no repetition avoidance involved), inpainting is performed with the prompt “a peripheral view of <dSCene> where we see <the corresponding description in d36o>”- If any object exists in drepeat, the positive prompt of “a peripheral view of <dScme> where we only see <the corresponding description in d360>” is used, and the negative prompt of “any type of <the object in drepeat>" is used (e.g., one sentence for each object in drepeat - The positive prompt template prevents stable diffusion from generating common objects of an environment (e.g., the bed in a bedroom). The negative prompt template avoids duplication of objects mentioned in drepeat-

[0033] Bias exists in the training data of diffusion models - an image with the caption of “a bedroom” mostly contains a bed. Therefore, repeated objects can still be generated even with constraints from the prompt. To further alleviate this problem, £v(-) is used to check whether each inpainted image Z(contains objects mentioned in drepeat (e.g., line 15). If the answer is “yes”, inpainting is re-run until the answer becomes “no” or the maximum number of trials c (e.g., twenty) is reached. As already noted, one or more prompts may be used to enable language models to automatically control diffusion models during inpainting. For line 2 of the pseudocode listing 40, the following two questions can be presented to the vision language model Lv(-):

[0034] QIBLIP Question: What is this place (describe with fewer than 5 words)? Answer:

[0035] Q2BLIP Question: Describe the foreground and background in detail and separately? Answer:

[0036] These two questions cause the model to output scene-level coarse and fine descriptions without focusing on centralized objects, which is beneficial for inferring the global scene layout at line 3. The final di is the answers of both questions.

[0037] To obtain scene layout descriptions dv,o of individual views, the following question is presented to the language model £(■) at line 3:

[0038] QIGPT Given a scene with <answer of Q1BLIP>, where in font of us we see <answer of Q2BLIP>. Generate 6 rotated views to describe what else you see in this place, where the camera of each view rotates 60 degrees to the right (you dont need to describe the original view, i.e., the first view of the 6 views you need to describe is the view with 60 degree rotation angle). Don’t involve redundant details, just describe the content of each view. Also don’t repeat the same object in different views. Don’t refer to previously generated views. Generate concise (<10 words) and diverse contents for each view. Each sentence starts with: View xxx(view number, from 1-6): We see...

[0039] Although the language model is asked to output results in the order of 60=>120..., the prompts can simply be used in the order of, for example, 60=> - 60=>120=>-120.... Moreover, the 60 degree value is merely a rough number so that the language model can easily understand the question (e.g., based on the actual angular rotation in the warp and inpaint operations, a value closer to 40 degrees is used rather than 60 degrees).

[0040] As already noted, the language model (e.g., ChatGPT) sometimes cannot fully follow the format request in QIGPT, which can make automatic prompt generation fail. To avoid such a failure, a check IS performed as to whether the output of QIGPT has the required number of lines (e.g., 6), and whether each line starts from ‘View XXX (line number): We see’ . Line 3 of the pseudo-code listing 40 is re-run if any of the conditions are violated. This approach ensures that the language model understands the question and satisfies all format requests. After all 360° perspective views are generated, the views are fused / merged into the final panorama. To ensure a smooth merging of multiple overlapping images into a panoramic image, and create a high resolution panorama, a smoothed fusion technique and super-resolution are leveraged.

[0041] More particularly, adjacent pixels at the center of an image have a larger angular distance than the pixels at the side of an image. When warping a completed view to a novel view, the original central region becomes the side region, making the rendered image blurry due to interpolation. Meanwhile, the panorama created by these images has a relatively low resolution because the resolution of the stable diffusion model output is 512*512. To address both problems, super-resolution is applied to the output It of each inpaint operation, increasing the resolution of to 2048 * 2048. Then, the high-resolution image is warped to a low-resolution novel view so that no (strong) interpolation is required. After performing all warp and inpaint operations, the superresolution images are simply fused to generate a high-resolution panorama.

[0042] During warping and panorama generation, multiple perspective images might have overlaps at the same region. To avoid sharp boundaries when merging the images, a weighted average is performed. For example, given multiple warped pixels at the same location with colors a, the final merged pixel color is cmerge= where the 4wi weight Wt is computed as the distance to the nearest image boundary at the original view i. This strategy effectively down- weights the pixels near the warping boundaries, ensuring a smooth transition during multi- view fusion.

[0043] To create the final panorama (e.g., line 22), each view is first projected to the unit sphere. Then, the equirectangular projection is performed to warp multiple projected views to the same equirectangular plane and the projected views are merged into a single equirectangular image.

[0044] To remove object-level information, line 4 of the pseudo-code listing 40 can ask:

[0045] Q2GPT Modify the sentence: <answer of Q1BLIP> SO that we remove all the objects from the description (e.g., “a bedroom with a bed” would become “a bedroom”. Do not change the sentence if the description is only an object). Just output the modified sentence.

[0046] To adaptively judge whether repeated objects are to be avoided, line 5 of the pseudo-code listing 40 can ask the following two questions: Q3GPT Given a scene with <answer of Q1BLIP>, where in font of us we see <answer of Q2BLIP>. What would be the two major foreground objects that we see? Use two lines to describe them where each line is in the format of “We see: xxx (one object, don’t describe details, just one word for the object. Start from the most possible object. Don’t mention background objects like things on the wall, ceiling or floor.)”

[0047] Q4GPT DO we often see multiple <each object in the answer of Q3GPT> in a scene with <answer of Q1BLIP> Just say ’yes’ or ’no’ with all lower case letters.

[0048] The final state of the drepeat variable is the set of objects in Q3GPT such that the corresponding answer of Q4GPT is “no”.

[0049] Quantitatively, embodiments consistently outperform conventional solutions in both human evaluations and the standard evaluation metric. The results of the technology described herein are preferred more frequently (e.g., greater than 50%) than the results from conventional solutions. Embodiments can also achieve a higher Inception Score (IS), indicating more diverse multi-view layouts and higher perspective rendering quality.

[0050] FIG. 3 shows comparative image- to-panorama visualization results. In the illustrated example, a first set of images 50 (50a, 50b, e.g., Stable Diffusion v2) demonstrates that the solution is unable to close the 360° loop (e.g., sharp boundary at the middle of the panorama). A second set of images 52 (52a, 52b, e.g., Text2room) and a third set of images 54 (54a, 54b, e.g., MVDiffusion) demonstrate that the solutions do not avoid duplicate objects in multiple views. By contrast, a fourth set of images 56 (56a, 56b, e.g., L-MAGIC) demonstrate that the technology described herein generates panoramas with high perspective rendering quality and reasonable scene layouts.

[0051] FIG. 4 shows a method 60 of operating a performance-enhanced computing system. The method 60 may be implemented in one or more modules as a set of logic instructions stored in a machine- or computer-readable storage medium such as random access memory (RAM), read only memory (ROM), programmable ROM (PROM), firmware, flash memory, etc., in hardware, or any combination thereof. For example, hardware implementations may include configurable logic, fixed- functionality logic, or any combination thereof. Examples of configurable logic (e.g., configurable hardware) include suitably configured programmable logic arrays (PLAs), field programmable gate arrays (FPGAs), complex programmable logic devices (CPLDs), and general purpose microprocessors. Examples of fixed- functionality logic (e.g., fixed- functionality hardware) include suitably configured application specific integrated circuits (ASICs), combinational logic circuits, and sequential logic circuits. The configurable or fixed-functionality logic can be implemented with complementary metal oxide semiconductor (CMOS) logic circuits, transistor-transistor logic (TTL) logic circuits, or other circuits.

[0052] Computer program code to carry out operations shown in the method 60 can be written in any combination of one or more programming languages, including an object-oriented programming language such as JAVA, SMALLTALK, C++ or the like and conventional procedural programming languages, such as the “C” programming language or similar programming languages. Additionally, logic instructions might include assembler instructions, instruction set architecture (ISA) instructions, machine instructions, machine dependent instructions, microcode, state-setting data, configuration data for integrated circuitry, state information that personalizes electronic circuitry and / or other structural components that are native to hardware (e.g., host processor, central processing unit / CPU, microcontroller, etc.).

[0053] Illustrated processing block 62 determines whether an input image is available from a camera output. If not, block 64 generates the input image based on a pre-trained diffusion model and one or more of a textual input, a sketch input or a synthetic image. Block 66 conducts a warp operation on the input image to obtain a shifted view and a mask, wherein the mask indicates missing regions in the shifted view associated with the warp operation. In one example, block 66 alternatively expands the field of view from both sides of the input image. Block 66 may also tune rotation angles to generate an incomplete region in the center of a final view. Block 68 provides for conducting, by a first pre-trained Al model (e.g., diffusion model), an inpaint operation on the missing regions based on a textual output from a second pre-trained Al model (e.g., language model), wherein the textual input is in response to one or more prompts. In one example, block 68 generates the prompt(s) based at least in part on a description of the input image from a third pre-trained Al model (e.g., vision language model). In an embodiment, the prompt(s) apply a distinction between scene layout regions (e.g., background, foreground). Additionally, the prompt(s) may include a negative prompt and a positive prompt. Moreover, block 68 may apply a super-resolution to an output of the inpaint operation.

[0054] A determination is made at block 70 as to whether a repeat of blocks 66 and 68 is to be conducted. In one example, the determination at block 70 takes into consideration whether the generated views are sufficient to complete a 360° panoramic scene. If a repeat of the warp operation and the inpaint operation is to be conducted, the method 60 returns to block 66 and a plurality of generated views are obtained. Otherwise, block 72 fuses the plurality of generated views into a 360° panoramic view. Additionally, block 74 may generate one or more of an immersive video or a 3D point cloud based on the 360° panoramic view.

[0055] To generate the video with pure camera rotations, image warping is used to project panoramas to individual frames based on the rotation matrix. To generate the video with camera translations, depth-based warping is used, where first a depth estimation model (e.g., Zoe-Depth or other arbitrary model) lifts pixels into 3D and the 3D point clouds are then warped into individual frames based on the camera extrinsic matrix (e.g., including rotation and translation).

[0056] To generate 3D point clouds, depth estimation models are first leveraged on individual perspective views of the panorama, and then the multi-view depth is optimized to force the overlapping pixels in multiple views to have the same depth. Then, the multi-view depth is lifted to 3D point clouds, which are fused into the same coordinate frame to produce the final 360-degree point clouds.

[0057] The method 60 therefore enhances performance at least to the extent that using the second pre-trained Al model to guide the inpaint operation results in higher quality images and more coherent scene layouts. Additionally, the ability to start from camera outputs (e.g., “in-the-wild” input images), synthetic images, sketch drawings of a perspective view and textual inputs renders the method 60 more extensible for a wider variety of applications. Indeed, the realistic and diverse scenes can be used to create 3D point clouds and immersive videos for applications such as video games, virtual reality (VR) and 3D environments.

[0058] Turning now to FIG. 5, a performance-enhanced computing system 280 is shown. The system 280 may generally be part of an electronic device / platform having computing functionality (e.g., personal digital assistant / PDA, notebook computer, tablet computer, convertible tablet, edge node, server, cloud computing infrastructure), communications functionality (e.g., smart phone), imaging functionality (e.g., camera, camcorder), media playing functionality (e.g., smart television / TV), wearable functionality (e.g., watch, eyewear, headwear, footwear, jewelry), vehicular functionality (e.g., car, truck, motorcycle), robotic functionality (e.g., autonomous robot), Internet of Things (loT) functionality, drone functionality, etc., or any combination thereof.

[0059] In the illustrated example, the system 280 includes a host processor 282 (e.g., central processing unit / CPU) having an integrated memory controller (IMC) 284 that is coupled to a system memory 286 (e.g., dual inline memory module / DIMM including dynamic RAM / DRAM). In an embodiment, an IO (input / output) module 288 is coupled to the host processor 282. The illustrated IO module 288 communicates with, for example, a display 290 (e.g., touch screen, liquid crystal display / LCD, light emitting diode / LED display), a camera 304, mass storage 302 (e.g., hard disk drive / HDD, optical disc, solid state drive / SSD) and a network controller 292 (e.g., wired and / or wireless). The host processor 282 may be combined with the IO module 288, a graphics processor 294, and an Al accelerator 296 (e.g., specialized processor) into a system on chip (SoC) 298.

[0060] In an embodiment, the Al accelerator 296 and / or the host processor 282 execute instructions 300 (e.g., executable program instructions) retrieved from the system memory 286 and / or the mass storage 302 to perform one or more aspects of the method 60 (FIG. 4), already discussed. Thus, execution of the instructions 300 causes the Al accelerator 296, the host processor 282 and / or the computing system 280 to conduct a warp operation on an input image to obtain a shifted view and a mask, wherein the mask indicates missing regions in the shifted view associated with the warp operation. In addition, execution of the instructions 300 causes the Al accelerator 296, the host processor 282 and / or the computing system 280 to conduct, by a first pre-trained Al model, an inpaint operation on the missing regions based on a textual output from a second pre-trained Al model, wherein the textual input is in response to one or more prompts. Execution of the instructions 300 causes the Al accelerator 296, the host processor 282 and / or the computing system 280 to repeat the warp operation and the inpaint operation to obtain a plurality of generated views, which may be fused into a 360° panoramic view.

[0061] The computing system 280 is therefore considered performance-enhanced at least to the extent that using the second pre-trained Al model to guide the inpainting operation results in higher quality images and more coherent scene layouts. Additionally, the ability to start from an output of the camera 304, synthetic images, sketch drawings of a perspective view and textual inputs renders the method 60 more extensible for a wide variety of applications. Indeed, the realistic and diverse scenes can be used to create 3D point clouds and immersive videos for applications such as video games, VR and 3D environments.

[0062] FIG. 6 shows a semiconductor apparatus 350 (e.g., chip, die, package). The illustrated apparatus 350 includes one or more substrates 352 (e.g., silicon, sapphire, gallium arsenide) and logic 354 (e.g., transistor array and other integrated circuit / IC components) coupled to the substrate(s) 352. In an embodiment, the logic 354 implements one or more aspects of the method 60 (FIG. 4), already discussed.

[0063] The logic 354 may be implemented at least partly in configurable or fixed- functionality hardware. In one example, the logic 354 includes transistor channel regions that are positioned (e.g., embedded) within the substrate(s) 352. Thus, the interface between the logic 354 and the substrate(s) 352 may not be an abrupt junction. The logic 354 may also be considered to include an epitaxial layer that is grown on an initial wafer of the substrate(s) 352.

[0064] FIG. 7 illustrates a processor core 400 according to one embodiment. The processor core 400 may be the core for any type of processor, such as a micro-processor, an embedded processor, a digital signal processor (DSP), a network processor, or other device to execute code. Although only one processor core 400 is illustrated in FIG. 7, a processing element may alternatively include more than one of the processor core 400 illustrated in FIG. 7. The processor core 400 may be a single-threaded core or, for at least one embodiment, the processor core 400 may be multithreaded in that it may include more than one hardware thread context (or “logical processor”) per core.

[0065] FIG. 7 also illustrates a memory 470 coupled to the processor core 400. The memory 470 may be any of a wide variety of memories (including various layers of memory hierarchy) as are known or otherwise available to those of skill in the art. The memory 470 may include one or more code 413 instruction(s) to be executed by the processor core 400, wherein the code 413 may implement the method 60 (FIG. 4), already discussed. The processor core 400 follows a program sequence of instructions indicated by the code 413. Each instruction may enter a front end portion 410 and be processed by one or more decoders 420. The decoder 420 may generate as its output a micro operation such as a fixed width micro operation in a predefined format, or may generate other instructions, microinstructions, or control signals which reflect the original code instruction. The illustrated front end portion 410 also includes register renaming logic 425 and scheduling logic 430, which generally allocate resources and queue the operation corresponding to the convert instruction for execution. The processor core 400 is shown including execution logic 450 having a set of execution units 455-1 through 455-N. Some embodiments may include a number of execution units dedicated to specific functions or sets of functions. Other embodiments may include only one execution unit or one execution unit that can perform a particular function. The illustrated execution logic 450 performs the operations specified by code instructions.

[0066] After completion of execution of the operations specified by the code instructions, back end logic 460 retires the instructions of the code 413. Tn one embodiment, the processor core 400 allows out of order execution but requires in order retirement of instructions. Retirement logic 465 may take a variety of forms as known to those of skill in the art (e.g., re-order buffers or the like). In this manner, the processor core 400 is transformed during execution of the code 413, at least in terms of the output generated by the decoder, the hardware registers and tables utilized by the register renaming logic 425, and any registers (not shown) modified by the execution logic 450.

[0067] Although not illustrated in FIG. 7, a processing element may include other elements on chip with the processor core 400. For example, a processing element may include memory control logic along with the processor core 400. The processing element may include I / O control logic and / or may include I / O control logic integrated with memory control logic. The processing element may also include one or more caches.

[0068] Referring now to FIG. 8, shown is a block diagram of a computing system 1000 embodiment in accordance with an embodiment. Shown in FIG. 8 is a multiprocessor system 1000 that includes a first processing element 1070 and a second processing element 1080. While two processing elements 1070 and 1080 are shown, it is to be understood that an embodiment of the system 1000 may also include only one such processing element.

[0069] The system 1000 is illustrated as a point-to-point interconnect system, wherein the first processing element 1070 and the second processing element 1080 are coupled via a point-to-point interconnect 1050. It should be understood that any or all of the interconnects illustrated in FIG. 8 may be implemented as a multi-drop bus rather than point-to-point interconnect.

[0070] As shown in FIG. 8, each of processing elements 1070 and 1080 may be multicore processors, including first and second processor cores (i.e., processor cores 1074a and 1074b and processor cores 1084a and 1084b). Such cores 1074a, 1074b, 1084a, 1084b may be configured to execute instruction code in a manner similar to that discussed above in connection with FIG. 7.

[0071] Each processing element 1070, 1080 may include at least one shared cache 1896a, 1896b. The shared cache 1896a, 1896b may store data (e.g., instructions) that are utilized by one or more components of the processor, such as the cores 1074a, 1074b and 1084a, 1084b, respectively. For example, the shared cache 1896a, 1896b may locally cache data stored in a memory 1032, 1034 for faster access by components of the processor. In one or more embodiments, the shared cache 1896a, 1896b may include one or more mid-level caches, such as level 2 (L2), level 3 (L3), level 4 (L4), or other levels of cache, a last level cache (LLC), and / or combinations thereof.

[0072] While shown with only two processing elements 1070, 1080, it is to be understood that the scope of the embodiments are not so limited. In other embodiments, one or more additional processing elements may be present in a given processor. Alternatively, one or more of processing elements 1070, 1080 may be an element other than a processor, such as an accelerator or a field programmable gate array. For example, additional processing element(s) may include additional processors(s) that are the same as a first processor 1070, additional processor(s) that are heterogeneous or asymmetric to processor a first processor 1070, accelerators (such as, e.g., graphics accelerators or digital signal processing (DSP) units), field programmable gate arrays, or any other processing element. There can be a variety of differences between the processing elements 1070, 1080 in terms of a spectrum of metrics of merit including architectural, micro architectural, thermal, power consumption characteristics, and the like. These differences may effectively manifest themselves as asymmetry and heterogeneity amongst the processing elements 1070, 1080. For at least one embodiment, the various processing elements 1070, 1080 may reside in the same die package.

[0073] The first processing element 1070 may further include memory controller logic (MC) 1072 and point-to-point (P-P) interfaces 1076 and 1078. Similarly, the second processing element 1080 may include a MC 1082 and P-P interfaces 1086 and 1088. As shown in FIG. 8, MC’s 1072 and 1082 couple the processors to respective memories, namely a memory 1032 and a memory 1034, which may be portions of main memory locally attached to the respective processors. While the MC 1072 and 1082 is illustrated as integrated into the processing elements 1070, 1080, for alternative embodiments the MC logic may be discrete logic outside the processing elements 1070, 1080 rather than integrated therein.

[0074] The first processing element 1070 and the second processing element 1080 may be coupled to an I / O subsystem 1090 via P-P interconnects 1076 1086, respectively. As shown in FIG. 8, the TO subsystem 1090 includes P-P interfaces 1094 and 1098. Furthermore, TO subsystem 1090 includes an interface 1092 to couple TO subsystem 1090 with a high performance graphics engine 1038. In one embodiment, bus 1049 may be used to couple the graphics engine 1038 to the I / O subsystem 1090. Alternately, a point-to-point interconnect may couple these components.

[0075] In turn, VO subsystem 1090 may be coupled to a first bus 1016 via an interface 1096. In one embodiment, the first bus 1016 may be a Peripheral Component Interconnect (PCI) bus, or a bus such as a PCI Express bus or another third generation VO interconnect bus, although the scope of the embodiments are not so limited.

[0076] As shown in FIG. 8, various VO devices 1014 (e.g., biometric scanners, speakers, cameras, sensors) may be coupled to the first bus 1016, along with a bus bridge 1018 which may couple the first bus 1016 to a second bus 1020. In one embodiment, the second bus 1020 may be a low pin count (LPC) bus. Various devices may be coupled to the second bus 1020 including, for example, a keyboard / mouse 1012, communication device(s) 1026, and a data storage unit 1019 such as a disk drive or other mass storage device which may include code 1030, in one embodiment. The illustrated code 1030 may implement the method 60 (FIG. 4), already discussed. Further, an audio VO 1024 may be coupled to second bus 1020 and a battery 1010 may supply power to the computing system 1000.

[0077] Note that other embodiments are contemplated. For example, instead of the point-to-point architecture of FIG. 8, a system may implement a multi-drop bus or another such communication topology. Also, the elements of FIG. 8 may alternatively be partitioned using more or fewer integrated chips than shown in FIG. 8.

[0078] In one example, the technology described herein is incorporated into the INTEL GAUDI Al accelerator platform to achieve multi-modal panorama generation Al functionality. The technology described herein may also be incorporated into the Al Personal Computer (PC) Acceleration Program, which connects independent hardware vendors (IHVs) and independent software vendors (ISVs) with INTEL resources including artificial intelligence (Al) toolchains, training, co-engineering, software optimization, hardware, design resources, technical expertise, co-marketing, and sales opportunities.

[0079] Additional Notes and Examples:

[0080] Example 1 includes a performance-enhanced computing system comprising a network controller, a processor coupled to the network controller, and a memory coupled to the processor, the memory including a plurality of executable program instructions, which when executed by the processor, cause the processor to conduct a warp operation on an input image to obtain a shifted view and a mask, wherein the mask indicates missing regions in the shifted view associated with the warp operation, conduct, by a first pre-trained artificial intelligence (Al) model, an inpaint operation on the missing regions based on a textual output from a second pre-trained Al model, wherein the textual output is in response to one or more prompts, and repeat the warp operation and the inpaint operation to obtain a plurality of generated views.

[0081] Example 2 includes the computing system of Example 1 , wherein the plurality of executable program instructions, when executed, further cause the processor to fuse the plurality of generated views into a 360° panoramic view.

[0082] Example 3 includes the computing system of Example 2, wherein the plurality of executable program instructions, when executed, further cause the processor to generate one or more of a video or a three-dimensional point cloud based on the 360° panoramic view.

[0083] Example 4 includes the computing system of any one of Examples 1 to 3, wherein the plurality of executable program instructions, when executed, further cause the processor to generate the one or more prompts based at least in part on a description of the input image from a third pre-trained Al model, wherein the one or more prompts are to apply a distinction between scene layout regions, and wherein the one or more prompts are to include a negative prompt and a positive prompt.

[0084] Example 5 includes the computing system of any one of Examples 1 to 3, wherein the one or more executable program instructions, when executed, further cause the processor to apply a super-resolution to an output of the inpaint operation.

[0085] Example 6 includes at least one computer readable storage medium comprising a plurality of executable program instructions, which when executed by a computing system, cause the computing system to conduct a warp operation on an input image to obtain a shifted view and a mask, wherein the mask indicates missing regions in the shifted view associated with the warp operation, conduct, by a first pre-trained artificial intelligence (Al) model, an inpaint operation on the missing regions based on a textual output from a second pre-trained Al model, wherein the textual output is in response to one or more prompts, and repeat the warp operation and the inpaint operation to obtain a plurality of generated views.

[0086] Example 7 includes the at least one computer readable storage medium of Example 6, wherein the plurality of executable program instructions, when executed, further cause the computing system to fuse the plurality of generated views into a 360° panoramic view.

[0087] Example 8 includes the at least one computer readable storage medium of Example 7, wherein the plurality of executable program instructions, when executed, further cause the computing system to generate one or more of a video or a three- dimensional point cloud based on the 360° panoramic view.

[0088] Example 9 includes the at least one computer readable storage medium of Example 6, wherein the plurality of executable program instructions, when executed, further cause the computing system to generate the one or more prompts based at least in part on a description of the input image from a third pre-trained Al model, wherein the one or more prompts are to apply a distinction between scene layout regions, and wherein the one or more prompts are to include a negative prompt and a positive prompt.

[0089] Example 10 includes the at least one computer readable storage medium of Example 6, wherein the one or more executable program instructions, when executed, further cause the computing system to apply a super-resolution to an output of the inpaint operation.

[0090] Example 11 includes the at least one computer readable storage medium of any one of Examples 6 to 10, wherein the one or more executable program instructions, when executed, further cause the computing system to generate the input image based on a pre-trained diffusion model and one or more of a textual input, a sketch input or a synthetic image.

[0091] Example 12 includes the at least one computer readable storage medium of any one of Examples 6 to 10, wherein the input image is to be associated with a camera output.

[0092] Example 13 includes a semiconductor apparatus comprising one or more substrates, and logic coupled to the one or more substrates, wherein the logic is implemented at least partly in one or more of configurable or fixed-functionality hardware, the logic to conduct a warp operation on an input image to obtain a shifted view and a mask, wherein the mask indicates missing regions in the shifted view associated with the warp operation, conduct, by a first pre-trained artificial intelligence (Al) model, an inpaint operation on the missing regions based on a textual output from a second pre-trained Al model, wherein the textual output is in response to one or more prompts, and repeat the warp operation and the inpaint operation to obtain a plurality of generated views.

[0093] Example 14 includes the semiconductor apparatus of Example 13, wherein the logic is further to fuse the plurality of generated views into a 360° panoramic view.

[0094] Example 15 includes the semiconductor apparatus of Example 14, wherein the logic is further to generate one or more of a video or a three-dimensional point cloud based on the 360° panoramic view.

[0095] Example 16 includes the semiconductor apparatus of Example 13, wherein the logic is to generate the one or more prompts based at least in part on a description of the input image from a third pre-trained Al model, wherein the one or more prompts are to apply a distinction between scene layout regions, and wherein the one or more prompts are to include a negative prompt and a positive prompt.

[0096] Example 17 includes the semiconductor apparatus of Example 13, wherein the logic is further to apply a super-resolution to an output of the inpaint operation.

[0097] Example 18 includes the semiconductor apparatus of any one of Examples 13 to 17, wherein the logic is further to generate the input image based on a pre- trained diffusion model and one or more of a textual input, a sketch input or a synthetic image.

[0098] Example 19 includes the semiconductor apparatus of any one of Examples 13 to 17, wherein the input image is to be associated with a camera output.

[0099] Example 20 includes the semiconductor apparatus of any one of Examples 13 to 19, wherein the logic coupled to the one or more substrates includes transistor regions that are positioned within the one or more substrates.

[0100] Example 21 includes a method of operating a performance-enhanced computing system, the method comprising conducting a warp operation on an input image to obtain a shifted view and a mask, wherein the mask indicates missing regions in the shifted view associated with the warp operation, conducting, by a first pre-trained artificial intelligence (Al) model, an inpaint operation on the missing regions based on a textual output from a second pre-trained Al model, wherein the textual output is in response to one or more prompts, repeating the warp operation and the inpaint operation to obtain a plurality of generated views.

[0101] Example 22 includes an apparatus comprising means for performing the method of Example 21.

[0102] Example sizes / models / values / ranges may have been given, although embodiments are not limited to the same. As manufacturing techniques (e.g., photolithography) mature over time, it is expected that devices of smaller size could be manufactured. In addition, well known power / ground connections to TC chips and other components may or may not be shown within the figures, for simplicity of illustration and discussion, and so as not to obscure certain aspects of the embodiments. Further, arrangements may be shown in block diagram form in order to avoid obscuring embodiments, and also in view of the fact that specifics with respect to implementation of such block diagram arrangements are highly dependent upon the computing system within which the embodiment is to be implemented, i.e., such specifics should be well within purview of one skilled in the art. Where specific details (e.g., circuits) are set forth in order to describe example embodiments, it should be apparent to one skilled in the art that embodiments can be practiced without, or with variation of, these specific details. The description is thus to be regarded as illustrative instead of limiting.

[0103] The term “coupled” may be used herein to refer to any type of relationship, direct or indirect, between the components in question, and may apply to electrical, mechanical, fluid, optical, electromagnetic, electromechanical or other connections. In addition, the terms “first”, “second”, etc. may be used herein only to facilitate discussion, and carry no particular temporal or chronological significance unless otherwise indicated.

[0104] As used in this application and in the claims, a list of items joined by the term “one or more of’ may mean any combination of the listed terms. For example, the phrases “one or more of A, B or C” may mean A; B; C; A and B; A and C; B and C; or A, B and C.

[0105] Those skilled in the art will appreciate from the foregoing description that the broad techniques of the embodiments can be implemented in a variety of forms. Therefore, while the embodiments have been described in connection with particular examples thereof, the true scope of the embodiments should not be so limited since other modifications will become apparent to the skilled practitioner upon a study of the drawings, specification, and following claims.

Claims

CLAIMSWe claim:

1. A performance-enhanced computing system comprising: a network controller; a processor coupled to the network controller; and a memory coupled to the processor, the memory including a plurality of executable program instructions, which when executed by the processor, cause the processor to: conduct a warp operation on an input image to obtain a shifted view and a mask, wherein the mask indicates missing regions in the shifted view associated with the warp operation, conduct, by a first pre-trained artificial intelligence (Al) model, an inpaint operation on the missing regions based on a textual output from a second pre-trained Al model, wherein the textual output is in response to one or more prompts, and repeat the warp operation and the inpaint operation to obtain a plurality of generated views.

2. The computing system of claim 1 , wherein the plurality of executable program instructions, when executed, further cause the processor to fuse the plurality of generated views into a 360° panoramic view.

3. The computing system of claim 2, wherein the plurality of executable program instructions, when executed, further cause the processor to generate one or more of a video or a three-dimensional point cloud based on the 360° panoramic view.

4. The computing system of any one of claims 1 to 3, wherein the plurality of executable program instructions, when executed, further cause the processor to generate the one or more prompts based at least in part on a description of the input image from a third pre-trained Al model, wherein the one or moreprompts are to apply a distinction between scene layout regions, and wherein the one or more prompts are to include a negative prompt and a positive prompt.

5. The computing system of any one of claims 1 to 3, wherein the one or more executable program instructions, when executed, further cause the processor to apply a super-resolution to an output of the inpaint operation.

6. At least one computer readable storage medium comprising a plurality of executable program instructions, which when executed by a computing system, cause the computing system to: conduct a warp operation on an input image to obtain a shifted view and a mask, wherein the mask indicates missing regions in the shifted view associated with the warp operation; conduct, by a first pre-trained artificial intelligence (Al) model, an inpaint operation on the missing regions based on a textual output from a second pre-trained Al model, wherein the textual output is in response to one or more prompts; and repeat the warp operation and the inpaint operation to obtain a plurality of generated views.

7. The at least one computer readable storage medium of claim 6, wherein the plurality of executable program instructions, when executed, further cause the computing system to fuse the plurality of generated views into a 360° panoramic view.

8. The at least one computer readable storage medium of claim 7, wherein the plurality of executable program instructions, when executed, further cause the computing system to generate one or more of a video or a three-dimensional point cloud based on the 360° panoramic view.

9. The at least one computer readable storage medium of claim 6, wherein the plurality of executable program instructions, when executed, further cause the computing system to generate the one or more prompts based at least in part on a description of the input image from a third pre-trained Al model, wherein the oneor more prompts are to apply a distinction between scene layout regions, and wherein the one or more prompts are to include a negative prompt and a positive prompt.

10. The at least one computer readable storage medium of claim 6, wherein the one or more executable program instructions, when executed, further cause the computing system to apply a super-resolution to an output of the inpaint operation.

11. The at least one computer readable storage medium of any one of claims 6 to 10, wherein the one or more executable program instructions, when executed, further cause the computing system to generate the input image based on a pre-trained diffusion model and one or more of a textual input, a sketch input or a synthetic image.

12. The at least one computer readable storage medium of any one of claims 6 to 10, wherein the input image is to be associated with a camera output.

13. A semiconductor apparatus comprising: one or more substrates; and logic coupled to the one or more substrates, wherein the logic is implemented at least partly in one or more of configurable or fixed-functionality hardware, the logic to: conduct a warp operation on an input image to obtain a shifted view and a mask, wherein the mask indicates missing regions in the shifted view associated with the warp operation; conduct, by a first pre-trained artificial intelligence (Al) model, an inpaint operation on the missing regions based on a textual output from a second pre-trained Al model, wherein the textual output is in response to one or more prompts; and repeat the warp operation and the inpaint operation to obtain a plurality of generated views.

14. The semiconductor apparatus of claim 13, wherein the logic is further to fuse the plurality of generated views into a 360° panoramic view.

15. The semiconductor apparatus of claim 14, wherein the logic is further to generate one or more of a video or a three-dimensional point cloud based on the 360° panoramic view.

16. The semiconductor apparatus of claim 13, wherein the logic is to generate the one or more prompts based at least in part on a description of the input image from a third pre-trained Al model, wherein the one or more prompts are to apply a distinction between scene layout regions, and wherein the one or more prompts are to include a negative prompt and a positive prompt.

17. The semiconductor apparatus of claim 13, wherein the logic is further to apply a super-resolution to an output of the inpaint operation.

18. The semiconductor apparatus of any one of claims 13 to 17, wherein the logic is further to generate the input image based on a pre- trained diffusion model and one or more of a textual input, a sketch input or a synthetic image.

19. The semiconductor apparatus of any one of claims 13 to 17, wherein the input image is to be associated with a camera output.

20. The semiconductor apparatus of any one of claims 13 to 17, wherein the logic coupled to the one or more substrates includes transistor regions that are positioned within the one or more substrates.

Images