Animation production method based on AI video transfer drawing

Abstract

The invention belongs to the technical field of artificial intelligence and computer image processing, and particularly relates to an animation production method based on AI video conversion, which comprises the following steps: acquiring a target video of a real person wearing a simplified cartoon doll image as an input material, decomposing a target video source file into a continuous frame sequence according to a preset frame rate, unifying the frame size through an Empty Latent Image node, and obtaining a frame sequence; outputting a to-be-processed image, constructing a ComfyUI node workflow to execute transplotting, generating inter-frame motion vector data, executing inter-frame consistency optimization, and synthesizing a processed frame sequence into an MP4 format animation video through a VHSVideo Combine node; the invention provides an end-to-end animation production method of ''lightweight role design + optical flow guide consistency optimization + adaptive sampling'', so as to realize cross-frame stability of the shape and style of a role. The method can adapt to various different scenes and dynamic changes, and can keep stable styles and expressions regardless of complex backgrounds or fast moving roles.

Description

TECHNICAL FIELD

[0001] The application relates to the field of artificial intelligence and computer image processing, and particularly relates to an animation production method based on AI video conversion. BACKGROUND

[0002] In recent years, diffusion models represented by Stable Diffusion and ControlNet have achieved remarkable results in single-frame image style transfer. However, when such models are directly applied to videos, due to frame-by-frame independent inference and lack of temporal constraints, phenomena such as "flickering", "jumping", and "structure drifting" easily occur, which seriously affect the visual experience of animations,

[0003] Currently, AI video conversion technology is widely used in video content creation. However, in actual applications, existing tools often face problems such as insufficient conversion consistency, style jumping, and unstable character performance. For example, the ControlNet node in ComfyUI and related workflows can achieve basic image conversion, but when processing videos, the animation style often appears inconsistent due to differences between frames, especially in complex scenes. Existing tools such as DomoAI, WarpVideoAI, and MagicAI also have similar problems, which leads to these tools not being widely used in animation production.

[0004] In recent years, diffusion models represented by Stable Diffusion and ControlNet have achieved remarkable results in single-frame image style transfer. However, when such models are directly applied to videos, due to frame-by-frame independent inference and lack of temporal constraints, phenomena such as "flickering", "jumping", and "structure drifting" easily occur, which seriously affect the visual experience of animations, and the existing technology has the following disadvantages:

[0005] Poor inter-frame consistency: Since existing technology mainly relies on frame-by-frame independent processing, the style and detail performance of each frame of video are inconsistent, especially in complex dynamic scenes, the jumping and jittering phenomenon between frames is serious, which affects the overall coherence and visual effect of the video.

[0006] High computational cost: Although diffusion models and other methods can produce high-quality single-frame images, they have high computational cost and long processing time, especially when processing high-resolution videos or long videos, the efficiency of actual application is low, which is difficult to meet the real-time needs of animation production.

[0007] Insufficient processing capability for complex scenes: Existing tools have difficulty maintaining the consistency of characters and scenes when processing rapidly changing or dynamically complex scenes, resulting in distorted conversion effects. For example, objects moving quickly or characters with rich details are prone to lose their original characteristics during the conversion process, and the final generated animation lacks stability and consistency.

[0008] Therefore we propose an animation production method based on AI video conversion to solve the above problems. SUMMARY

[0009] (I) Technical problems solved

[0010] In view of the deficiencies in the prior art, the present application provides an animation production method based on AI video conversion, aiming to overcome the problems of inter-frame inconsistency, high computational cost, and complex scene distortion in the prior art, and to provide an end-to-end animation production method of "lightweight character design + optical flow guided consistency optimization + adaptive sampling" to realize cross-frame stability of character appearance and style; significantly reduce the number of diffusion model iteration steps; and have robustness to fast motion and complex background, solving the problems proposed in the above background art.

[0011] (II) Technical solutions

[0012] In order to achieve the above purpose, the present application specifically adopts the following technical solutions:

[0013] An animation production method based on AI video conversion, comprising the following steps:

[0014] S1, obtaining a target video of a real person wearing a simplified cartoon doll image as input material, the cartoon doll image having geometric features of simple structure and smooth surface;

[0015] S2, decomposing the target video source file into a continuous frame sequence according to a preset frame rate, and performing resolution uniform processing on each frame to generate a standardized frame sequence;

[0016] S3, uniformly sizing the frames through an Empty Latent Image node, and outputting a to-be-processed image;

[0017] S4, constructing a ComfyUI node workflow to perform conversion;

[0018] S5, calculating the optical flow field of adjacent frames in the initial stylized frame sequence using an OpticalFlow node to generate inter-frame motion vector data;

[0019] S6, performing inter-frame consistency optimization, recombining latent images through a RebatchLatents node based on the motion vector data, and compensating for inter-frame differences using a smoothing filter algorithm to generate a consistent optimized latent frame sequence;

[0020] S7, combining the processed frame sequence into an MP4 format animation video through a VHS_VideoCombine node.

[0021] Further, the cartoon doll image satisfies:

[0022] The head adopts a spherical or ellipsoidal design, occupying 1 / 3 of the overall proportion of the character;

[0023] The body adopts a short and wide cylindrical structure, and the limbs adopt a cylinder with an aspect ratio of 3:1;

[0024] The palm adopts an arc-shaped glove design without knuckles;

[0025] The color scheme includes 1-2 main tones with lightness value > 70 LAB and contrast point colors with lightness difference > 50 LAB.

[0026] Further, in step S2, the target video is decomposed into a frame sequence by the VHS_LoadVideo node, and the parameter analysis includes: frame rate locking 24fps; resolution keeping 1080p; maximum frame loading number ≤1000.

[0027] Further, in step S4, the workflow is constructed to perform the redrawing: the control_vllp_sd15_linear_fp16.safetensors style model is loaded by the ControlNetLoaderAdvanced node;

[0028] The style migration is performed frame by frame by the ControlNetApplyAdvanced node, and the model application intensity coefficient is set to 0.65-0.85.

[0029] Further, the inter-frame consistency optimization in step S6 includes:

[0030] The optical flow motion vector of adjacent frames is calculated by the OpticalFlow node;

[0031] The latent image is reorganized based on the optical flow vector by the RebatchLatents node;

[0032] The KSampler node is used to perform 8-step sampling in SGM_uniform mode.

[0033] Further, the inter-frame consistency of step S6 further includes: establishing a color gradient constraint equation:

[0034]

[0035] Where β = 0.75-0.95 is the color attenuation factor, and γ = 0.8-1.2 is the optical flow weight coefficient;

[0036] A Gaussian smoothing filter with a radius of 3-5 pixels is applied to the object boundary.

[0037] Further, the inter-frame consistency of step S6 further includes:

[0038] Linear interpolation is performed on the potential vectors of adjacent frames, and the interpolation weight α is 0.3-0.7.

[0039] Mask protection is performed on the area where the motion vector exceeds the threshold value, wherein the threshold value is 10 pixels, to avoid over-smoothing and loss of details.

[0040] Further, the video synthesis stage of the step S7 comprises: adding an FFmpeg-based timestamp synchronization module, and setting the time base to 1 / 1000 second; enabling an H.264 encoder, and setting the CRF value to 18-22; and inserting an I-frame interval of 24 frames of GOP structure.

[0041] Further, the dynamic adaptability design of the cartoon figure image comprises:

[0042] The joint part adopts a spherical connection structure, and the rotation degree of freedom is ≤30°.

[0043] The anti-deformation reinforcement area is set in the fast moving part, and the deformation tolerance is <5%.

[0044] The high light area adopts a fixed color coordinate, x=0.45±0.05, and y=0.40±0.05.

[0045] Further, the method further comprises: after the video synthesis step, encoding the output video by FFmpeg to H.264, and setting the code rate to 8-12 Mbps; and performing color space conversion on the encoded video to convert RGB to YUV420p format to be compatible with mainstream players.

[0046] (III) Beneficial effects

[0047] Compared with the prior art, the present application provides an animation production method based on AI video conversion, which has the following beneficial effects:

[0048] Simplified character design improves consistency: the present application adopts simplified geometric shapes and color matching strategies in character design; the character modeling structure is simple and the surface is smooth, reducing the style changes caused by the complexity of details in the conversion process; this simplified design not only reduces the processing difficulty, but also significantly improves the shape consistency and visual coherence of the animation characters in inter-frame conversion, ensuring that the performance of the characters in the animation is more stable.

[0049] Improve inter-frame consistency: by inputting the video of the figure image, the problem of inconsistent inter-frame style in the prior art is solved, and the generated video style is coherent, reducing the frame skipping and jitter phenomenon in the animation.

[0050] Adapting to diverse animation scenarios: Through adaptive and consistent character design and style management, this invention can adapt to a variety of different scenarios and dynamic changes, maintaining a stable style expression regardless of complex backgrounds or fast-moving characters. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating an AI-based video-to-image animation production method according to the present invention.

[0052] Figure 2 This is a schematic diagram showing the comparison of cartoon character redrawing using an AI video redrawing animation production method according to the present invention (the left side is the original image, and the right side is the redrawn image). Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] like Figure 1 and Figure 2 As shown, an embodiment of the present invention proposes an animation production method based on AI video transcribing, which includes the following steps:

[0055] S1. Obtain a target video of a real person wearing a simplified cartoon mascot image as input material. The cartoon mascot image has simple structure and smooth surface geometric features.

[0056] S2. Decompose the target video source file into a continuous frame sequence according to a preset frame rate, and perform resolution unification processing on each frame to generate a standardized frame sequence.

[0057] S3. Unify the frame size using the Empty Latent Image node and output the image to be processed;

[0058] S4. Construct a ComfyUI node workflow and execute the redraw;

[0059] S5. Calculate the optical flow field of adjacent frames in the initial stylized frame sequence using the OpticalFlow node to generate inter-frame motion vector data;

[0060] S6. Perform inter-frame consistency optimization. Based on the motion vector data, reconstruct the potential image through the RebatchLatents node, and use a smoothing filtering algorithm to compensate for inter-frame differences, generating a consistency-optimized potential frame sequence.

[0061] S7, combine the processed frame sequence into an MP4 format animation video through the VHS_VideoCombine node.

[0062] The application simplifies the character design to improve consistency: the application adopts simplified geometric shapes and color matching strategies in character design; the character modeling is simple in structure and smooth in surface, reducing the style changes caused by the complexity of details in the conversion process; this simplified design not only reduces the processing difficulty, but also significantly improves the shape consistency and visual coherence of the animation character in the interframe conversion, ensuring that the performance of the character in the animation is more stable. Improve interframe consistency: through the video input of the doll image, the problem of inconsistent interframe style in the prior art is solved, and the generated video style is coherent, reducing the frame skipping and jitter phenomenon in the animation. Adapt to various animation scenes: through adaptive consistent character design and style management, the application can adapt to various different scenes and dynamic changes, whether it is a complex background or a fast-moving character, it can maintain stable style performance.

[0063] As shown in Figure 1 and Figure 2 In some embodiments, the cartoon doll image meets: the head adopts a spherical or ellipsoidal design, and the volume accounts for 1 / 3 of the overall proportion of the character; the body adopts a short and wide cylindrical structure, and the limbs adopt a cylinder with an aspect ratio of 3:1; the palm adopts an arc-shaped glove design without knuckles; the color scheme includes 1-2 main tones with a lightness value >70LAB and contrast point colors with a lightness difference >50LAB.

[0064] The design specification of the cartoon doll image in step S1 includes:

[0065] a) The ratio of the head spherical curvature radius to the limb cylindrical diameter is 1.2:1;

[0066] b) The aspect ratio of the glove-shaped palm is 1.2:1, and the edge adopts a circular arc transition with a radius of 2-3mm;

[0067] c) The abdominal region is provided with a contrast color block with a lightness difference >30LAB from the main tone.

[0068] As shown in Figure 1 and Figure 2 In some embodiments, the target video is decomposed into a frame sequence through the VHS_LoadVideo node in step S2, and the analysis parameters include: frame rate locking 24fps; resolution remains 1080p; maximum frame loading number ≤1000.

[0069] The step S2 includes: performing CLAHE contrast enhancement on low-illumination frames, with a threshold limit of 40-60lux; using the Canny edge detection algorithm to extract feature lines, with a gradient threshold set to 50-150.

[0070] The workflow execution in step S4 constructs the transfer drawing: the control_vllp_sd15_linear_fp16.safetensors style model is loaded by the ControlNetLoaderAdvanced node;

[0071] The style migration is executed frame by frame by the ControlNetApplyAdvanced node, and the model application intensity coefficient is set to 0.65-0.85.

[0072] The ComfyUI workflow is executed in a hardware environment equipped with:

[0073] GPU: NVIDIA RTX 4090D graphics card, video memory ≥24GB;

[0074] Video memory bandwidth ≥1TB / s, single-precision floating-point performance ≥100TFLOPS;

[0075] Memory ≥64GB DDR5, memory bandwidth ≥500GB / s.

[0076] The parameter configuration in step S4 includes: setting the style fusion intensity δ=0.7±0.15 and the line retention coefficient λ=0.85±0.1 in the ControlNetApplyAdvanced node; and setting the detail retention threshold θ=0.6±0.1.

[0077] The ComfyUI workflow further includes: adjusting the input video resolution to a fixed size of 512×512 or 768×768 by the Empty Latent Image node; and setting the control weight to 0.8-1.2 in the ControlNetApplyAdvanced node to balance the style intensity and structure retention.

[0078] As shown in Figure 1 and Figure 2 In some embodiments, the inter-frame consistency optimization in step S6 includes:

[0079] The optical flow motion vector of adjacent frames is calculated by the OpticalFlow node;

[0080] The latent image is reorganized based on the optical flow vector by the RebatchLatents node;

[0081] The KSampler node is used for 8-step sampling in SGM_uniform mode.

[0082] The inter-frame consistency of step S6 further includes: establishing a color gradient constraint equation:

[0083]

[0084] wherein β = 0.75-0.95 is a color attenuation factor, and γ = 0.8-1.2 is a light flow weight coefficient;

[0085] Applying a Gaussian smoothing filter with a radius of 3-5 pixels to the object boundary.

[0086] The inter-frame consistency of the step S6 further comprises:

[0087] Linearly interpolating the latent vectors of adjacent frames with an interpolation weight a = 0.3-0.7;

[0088] Masking the regions where the motion vector exceeds a threshold value, wherein the threshold value = 10 pixels, to avoid over-smoothing and loss of details.

[0089] As shown in Figure 1 and Figure 2 In some embodiments, the video synthesis stage of the step S7 comprises: adding an FFmpeg-based timestamp synchronization module, with a time base set to 1 / 1000 seconds; enabling an H.264 encoder, with a CRF value set to 18-22; and inserting an I-frame interval of 24 frames of GOP structure.

[0090] As shown in Figure 1 and Figure 2 In some embodiments, the dynamic adaptability design of the cartoon figurine image comprises:

[0091] The joint positions adopt a spherical connection structure, with a rotational freedom degree ≤ 30°;

[0092] The fast-moving positions are provided with an anti-deformation reinforcement area, with a deformation tolerance < 5%;

[0093] The highlight areas adopt fixed color coordinates, with x = 0.45 ± 0.05 and y = 0.40 ± 0.05.

[0094] The color design of the cartoon figurine image satisfies: the main tone is a single high-saturation color, and the auxiliary color is a contrast color and does not exceed two; the color boundary adopts a hard edge design to avoid gradual transition.

[0095] As shown in Figure 1 and Figure 2 In some embodiments, the method further comprises: after the video synthesis step, encoding the output video through FFmpeg in H.264, with a code rate set to 8-12 Mbps; and performing color space conversion on the encoded video, converting RGB to YUV420p format to be compatible with mainstream players.

[0096] A system for performing the above method, comprising: a video acquisition unit (1080p camera); a preprocessing unit (CPU + 64GB RAM); a style transfer unit (GPU, 24GB video memory); an optical flow calculation unit (RAFT-small, TensorRT acceleration); a sampling and post-processing unit (shared GPU); an output unit (MP4 packer).

[0097] In use, the steps are described as follows:

[0098] Step S1: Character / scene lightweight design, geometric constraints: head adopts a sphere, accounting for 1 / 3 of the whole body; body and limbs are cylindrical, with a sphere transition at the joints; cancel the details of fingers / toes, and model them in a "glove / sock" integrated manner; material constraints: use a single matte material, roughness 0.3-0.5, no texture mapping; color constraints: 1-2 main colors, ≤2 secondary colors, lightness difference ≥40%, saturation ≥70%, to ensure consistency across lighting. This design minimizes high-frequency detail signals, providing a "low-entropy" input for the subsequent diffusion model, and significantly suppressing style jumps.

[0099] Step S2: Video acquisition and preprocessing, input: 1080p / 24fps MP4, YUV420, 8-bit; denoising: use BM3D temporal filtering, σ=5; frame alignment: if there is camera shake, use L1-Optical-Flow for global motion compensation; size normalization: uniformly scale to 512x512 or 768x768 through the EmptyLatentImage node, fill with black borders while maintaining the aspect ratio, and avoid distortion.

[0100] Step S3: Initial style transfer,

[0101] Model: ControlNetv1.1lineart_fp16, weight 1.0; Prompt template: "a cartoon character, simple clean lineart, flat color, no shadow";

[0102] Negative Prompt: "photo, realistic, texture, wrinkle"; CFGScale: 7-12; iteration steps: 8 (after experiments, 8 steps can reduce 40% of the calculation while maintaining quality);

[0103] Each frame is independently inferred to obtain a primary stylized latent variable Z0.

[0104] Step S4: Optical flow consistency optimization:

[0105] S4.1 Optical flow estimation, RAFT-small is adopted to calculate bidirectional optical flow F at 1 / 4 resolution t → t+1 , F t+1 → t ; Gaussian filter (σ = 1.5) and threshold cut-off (max displacement 20px) are performed on the optical flow to remove outliers.

[0106] S4.2 Potential space warping, F t → t+1 is used to perform differentiable bilinear warping on Z0 to obtain warped latent variable Z' t+1 ; Consistency loss is calculated: L_consist = ||Z' t+1 - Z 0t+1 ||2 2 ;

[0107] Update rule: Z 1t+1 = α·Z' t+1 + (1-α)·Z 0t+1 , α ∈ [0.3, 0.7], the larger the motion, the larger α.

[0108] 4.3 Edge protection, Sobel edge detection is used to generate a mask M_edge for the character contour edge; the value of α is reduced to 0.1 in the M_edge region to prevent excessive smoothing of the edge.

[0109] Step S5: Adaptive sampling and post-processing, KSampler (Euler, SGM_uniform) is used for the optimized latent variable Z1 with 8 steps; high-resolution repair: Latent Upscale + Img2Img secondary sampling is used with a magnification of 2x and 4 resampling steps; color matching: Histogram Matching is used to transfer the color distribution of the first frame to the remaining frames, with a global tonal difference ΔE < 2.

[0110] Step S6: Video synthesis and encoding, write frame by frame through the VHS_VideoCombine node;

[0111] Encoding parameters: H.264High@L4.2, CRF = 18, preset = slow, tune = animation;

[0112] Audio: If the source video contains audio tracks, use AAC 128kbps to synchronize and package.

[0113] Cartoon character design: In the AI cartoon image video conversion consistency optimization method of the present invention, the simplification and reasonable planning of character design is the key step to ensure the consistency of video conversion effect. Through the induction of multiple cartoon character design cases, it is found that simplifying geometric shapes, color matching and proportion design play an important role in animation production. This embodiment further describes how to apply these design characteristics to the consistency optimization of video conversion by summarizing the design characteristics of these characters.

[0114] Role modeling design and structure: Simplified character design not only helps to reduce the loss of details in the conversion process, but also improves the inter-frame consistency and ensures the stable performance of the character in different dynamic scenes. Based on cartoon character design cases, the main modeling characteristics of the character include: head design: the head adopts a round spherical design with a large volume, occupying one-third of the overall proportion of the character. This exaggerated head design enhances the affinity and cuteness of the character, and is easy to identify and maintain shape consistency during the conversion process. For example, the head design in Teletubbies, the round head with simple shape makes the character maintain good conversion effect even in fast movement or complex scenes. Body and limb design: the body of the character is a short and wide cylinder, avoiding complex muscles or bones, and the details of fingers or toes, so that the structure performance can be maintained stable during conversion. The limbs are designed as simple cylinders, and the palms are designed as "gloves", simplifying the details of the fingers. This design helps to reduce the complexity during animation movement.

[0115] Color design and matching: Color is an important part of character design, which not only affects visual performance, but also maintains consistency and coherence of style during conversion. The present invention adopts strong contrast, simple and clear color matching in character color design: main color: the color design of cartoon characters usually adopts bright single main color, supplemented by high-contrast auxiliary color. For example, the green and purple of Teletubbies, and the yellow and red matching of the chicken-shaped doll, make the character maintain prominent visual effect in the animation background. Contrast color: the red feather-like decoration on the head of the character and the red blanket in the hand form a strong contrast, such as the red crest on the head of the chicken-shaped doll, which forms a sharp contrast with the overall yellow color, helping to maintain the recognition of the character in complex scenes. Simplification and uniformity of color: the character usually adopts 1-2 main colors, supplemented by a small amount of contrast color to highlight important visual elements. This design simplifies the performance of the character under different light conditions, ensuring that the color does not deviate during conversion. For example, the design of pink cat, the whole body is mainly pink, supplemented by white belly, effectively improving the color consistency during conversion.

[0116] Cartoon image proportion design, the proportion design of the role follows the common characteristics of cartoon characters, the head-body proportion is 3:1, the larger head makes the expression more prominent, and at different angles and dynamics, the facial features can be ensured to be consistent in the conversion process. Head-body proportion: the head of the role accounts for one-third of the overall proportion, such exaggerated proportion design conforms to the style of cartoon characters, and due to the obvious head features, the morphological deformation caused by angle change is reduced during video conversion. Limb proportion: the whole role is based on circular and cylindrical geometry, reducing the error caused by complex details. This design simplifies the shape calculation and style transfer during conversion, so that better morphological consistency can be maintained in complex dynamic scenes.

[0117] The present application has the following advantages:

[0118] Simplified role design improves consistency: the present application adopts simplified geometric shapes and color matching strategies in role design; the role modeling structure is simple and the surface is smooth, reducing the style changes caused by complex details in the conversion process; this simplified design not only reduces the processing difficulty, but also significantly improves the morphological consistency and visual coherence of the animation role in inter-frame conversion, ensuring that the performance of the role in the animation is more stable.

[0119] Improve inter-frame consistency: by inputting the video of the doll image, the problem of inconsistent inter-frame style in the prior art is solved, and the generated video style is coherent, reducing the frame skipping and jitter phenomenon in animation.

[0120] Adapt to diverse animation scenes: through adaptive consistent role design and style management, the present application can adapt to various different scenes and dynamic changes, whether it is a complex background or a fast-moving role, it can maintain stable style performance.

[0121] Finally, it should be noted that: the above only describes the preferred embodiments of the present application and is not intended to limit the present application, although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent replacements for some technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An animation production method based on AI video transfer, characterized in that, Includes the following steps: S1. Obtain a target video of a real person wearing a simplified cartoon mascot image as input material. The cartoon mascot image has simple structure and smooth surface geometric features. S2. Decompose the target video source file into a continuous frame sequence according to a preset frame rate, and perform resolution unification processing on each frame to generate a standardized frame sequence. S3. Unify the frame size using the Empty Latent Image node and output the image to be processed; S4. Construct a ComfyUI node workflow and execute the redraw; S5. Calculate the optical flow field of adjacent frames in the initial stylized frame sequence using the OpticalFlow node to generate inter-frame motion vector data; S6. Perform inter-frame consistency optimization. Based on the motion vector data, reconstruct the potential image through the RebatchLatents node, and use a smoothing filtering algorithm to compensate for inter-frame differences, generating a consistency-optimized potential frame sequence. S7. Use the VHS_VideoCombine node to synthesize the processed frame sequence into an MP4 format animation video.

2. The animation production method based on AI video transfer according to claim 1, characterized in that: The cartoon doll image meets the following requirements: The head is designed as a sphere or ellipsoid, and its volume accounts for 1 / 3 of the overall size of the character; The body adopts a short and wide cylindrical structure, and the limbs adopt a cylindrical structure with a length-to-diameter ratio of 3:1; The palm features a curved, glove-like design without knuckles; The color scheme includes 1-2 primary colors with a brightness value > 70 LAB and contrasting accent colors with a brightness difference > 50 LAB.

3. The animation production method based on AI video transfer according to claim 1, characterized in that: In step S2, the target video is decomposed into a frame sequence through the VHS_LoadVideo node. The parsing parameters include: frame rate locked at 24fps; resolution maintained at 1080p; maximum number of frames loaded ≤ 1000.

4. The animation production method based on AI video transfer according to claim 1, characterized in that: In step S4, the workflow is constructed to perform the mapping: the control_vllp_sd15_linear_fp16.safetensors style model is loaded by the ControlNetLoaderAdvanced node; Style transfer was performed frame-by-frame using the ControlNetApplyAdvanced node, with the model application strength coefficient set to 0.65-0.

85.

5. The animation production method based on AI video transfer according to claim 1, characterized in that: The inter-frame consistency optimization performed in step S6 includes: Calculate the optical flow motion vectors of adjacent frames using OpticalFlow nodes; Reconstruct the latent image based on optical flow vectors using the RebatchLatents node; The KSampler node is used to perform 8-step sampling in SGM_uniform mode.

6. The animation production method based on AI video transfer according to claim 1, characterized in that: The inter-frame consistency step S6 further includes: establishing a color gradient constraint equation: Where β = 0.75-0.95 is the color attenuation factor, and γ = 0.8-1.2 is the optical flow weighting coefficient; Apply a Gaussian smoothing filter with a radius of 3-5 pixels to the object boundary.

7. The animation production method based on AI video transfer according to claim 1, characterized in that: The inter-frame consistency step S6 also includes: Linear interpolation is performed on the latent vectors of adjacent frames, with interpolation weights α = 0.3-0.7; For regions where the motion vector exceeds a threshold of 10 pixels, a mask is applied to protect the area and prevent excessive smoothing from causing loss of detail.

8. The animation production method based on AI video transfer according to claim 1, characterized in that: The video compositing stage in step S7 includes: adding an FFmpeg-based timestamp synchronization module with a time base set to 1 / 1000 second; enabling an H.264 encoder with a CRF value set to 18-22; and inserting a GOP structure with an I-frame interval of 24 frames.

9. The animation production method based on AI video transfer according to claim 1, characterized in that: The dynamic adaptive design of the cartoon mascot includes: The joints use a spherical connection structure with a rotational degree of freedom of ≤30°. Deformation-resistant reinforcement zones are set in rapidly moving parts, with a deformation tolerance of <5%; The highlight areas use fixed color coordinates: x = 0.45 ± 0.05, y = 0.40 ± 0.

05.

10. The animation production method based on AI video transfer according to claim 1, characterized in that: The method further includes: after the video compositing step, encoding the output video using FFmpeg with H.264 and setting the bitrate to 8-12Mbps; and performing color space conversion on the encoded video, converting RGB to YUV420p format to be compatible with mainstream players.

Images