Research map / Image generation / Text-to-image diffusion
Make-It-Vivid: Dressing Your Animatable Biped Cartoon Characters from Text
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Make-It-Vivid fine-tunes a LoRA-adapted Stable Diffusion 1.5 on VQA-captioned 3DBiCar UV texture maps and adds an adversarial loss against ControlNet depth-guided renderings to generate text-driven UV textures for biped cartoon characters.
Problem
Generating high-quality, text-conditioned UV textures for 3D biped cartoon characters, where high-quality paired text-texture data is scarce and existing texturing methods yield blurry details and seam artifacts.
Approach
Customize a pretrained text-to-image diffusion model with LoRA on UV texture maps for template-structured generation, and introduce adversarial training against ControlNet-synthesized multi-view images as a proxy for high-frequency realistic detail, with VQA-agent (BLIP2 + ChatGPT) captioning to build paired data.
Key results
- Achieves CLIP score 29.86 on the 300-prompt test set, versus 29.25 for TEXTure and 29.20 for Fantasia3D.
- Reaches FID 35.25 and KID 5.25 (x10^-3) on 3DBiCar versus 42.55 and 6.37 for Rabit.
- Generates a texture in about 0.03 minutes (about 2 seconds) on a single GPU, versus 2.38 minutes for TEXTure.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Make-It-Vivid instead of the strongest listed baseline | Text-guided UV texture generation · Test prompt set | CLIP: 29.25 -> 29.86 | Helped |
| use Make-It-Vivid instead of the strongest listed baseline | Text-guided UV texture generation · 3DBiCar | FID: 42.55 -> 35.25 | Helped |
| use Make-It-Vivid instead of the strongest listed baseline | Text-guided UV texture generation · 3DBiCar | KID(×10−3): 6.37 -> 5.25 | Helped |
| use LoRA rank 64 instead of rank 8 (ablation) | Make-It-Vivid · 3DBiCar texture generation | texture quality | Hurt |
| reduce LoRA rank from 64 to 8 (ablation) | Make-It-Vivid · 3DBiCar texture generation | texture quality | Mixed |
| add adversarial training (ablation) | Make-It-Vivid with LoRA rank 8 · 3DBiCar texture generation | fine-grained texture details | Helped |
Surprising findings
- Large LoRA rank (64) introduces sawtooth artifacts, while rank 8 gives more plausible semantic details.
- Simple LDM fine-tuning yields structure without fine-grained details such as cloth wrinkles, motivating the adversarial step.
Limitations
- Adversarial supervision relies on ControlNet-generated images as a proxy rather than real high-frequency UV data.
- Seam artifacts are handled with heuristic Gaussian filtering and image restoration rather than a principled solution.
- Evaluation relies largely on CLIP score, FID/KID and a user study, with a single template-based 300-prompt test set.
Experimental setup
Datasets: 3DBiCar ControlNet-synthesized multi-view images
Benchmarks: 300-prompt text-guided texture test set 3DBiCar dataset renders (FID/KID)
Metrics: CLIP score FID KID inference time GPU memory user study
Baselines: Rabit LatentPaint Fantasia3D Text2Tex TEXTure
Builds on: Stable Diffusion LoRA Rabit ControlNet BLIP-2
Closest papers by meaning
- Make-It-Animatable: An Efficient Framework for Authoring Animation-Ready 3D Characters (CVPR 2025 · similarity 0.85)
- Paint-it: Text-to-Texture Synthesis via Deep Convolutional Texture Map Optimization and Physically-Based Rendering (CVPR 2024 · similarity 0.85)
- DRiVE: Diffusion-based Rigging Empowers Generation of Versatile and Expressive Characters (CVPR 2025 · similarity 0.84)
- StdGEN: Semantic-Decomposed 3D Character Generation from Single Images (CVPR 2025 · similarity 0.84)
- X-Oscar: A Progressive Framework for High-quality Text-guided 3D Animatable Avatar Generation (ICML 2024 · similarity 0.83)
- DreamText: High Fidelity Scene Text Synthesis (CVPR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.