Research map / Image generation / Editing and style transfer
Uncovering the Disentanglement Capability in Text-to-Image Diffusion Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Partial text-embedding replacement (neutral to styled prompt only at later denoising steps) with fixed noise, instead of full replacement | Stable diffusion v1-4 · Qualitative example (person smiling) | identity preservation and attribute change | Helped |
| Optimize soft mixing weights of two text embeddings (about 50 parameters) with CLIP directional loss and perceptual loss, without fine-tuning the diffusion model | Stable diffusion v1-4 image editing · Human evaluation vs DiffusionCLIP (Celeb-A, LSUN-church) | user preference on attribute similarity, content preservation, overall quality | Helped |
| Use the proposed method on sketch and pixar human edits instead of DiffusionCLIP | Human-related editing · Human evaluation vs DiffusionCLIP | user preference | Hurt |
| Optimize mixing weights on 5 images instead of 1 or 3 (ablation) | Proposed method · Transfer to unseen images (qualitative) | identity preservation | Helped |
| Vary the way the target attribute is appended to c(1) (comma, direct, with) (ablation) | Proposed method · Image editing qualitative robustness | edit result consistency | No clear effect |
| Use a captioning-model generated description as c(0) instead of short or long handwritten description (ablation) | Proposed method · Image editing qualitative robustness | edit strength | Hurt |
| Use a longer c(1) with more correlated modifiers for the target attribute (ablation) | Proposed method · Image editing qualitative robustness | edit strength | Helped |
| Use Blended Diffusion/Imagic/Prompt-to-Prompt comparisons (qualitative) | Proposed method · Qualitative comparison with other diffusion editing baselines | content preservation | Mixed |
| Attempt to disentangle small accessory edits (hats, earrings, cake toppings) | Proposed method · Attribute disentanglement exploration | attribute disentanglement success | Hurt |
Closest papers by meaning
- NoiseCLR: A Contrastive Learning Approach for Unsupervised Discovery of Interpretable Directions in Diffusion Models (CVPR 2024 · similarity 0.90)
- E$^2$GAN: Efficient Training of Efficient GANs for Image-to-Image Translation (ICML 2024 · similarity 0.89)
- You Only Sample Once: Taming One-Step Text-to-Image Synthesis by Self-Cooperative Diffusion GANs (ICLR 2025 · similarity 0.88)
- Adversarial Supervision Makes Layout-to-Image Diffusion Models Thrive (ICLR 2024 · similarity 0.88)
- When StyleGAN Meets Stable Diffusion: a W+ Adapter for Personalized Image Generation (CVPR 2024 · similarity 0.87)
- DisenBooth: Identity-Preserving Disentangled Tuning for Subject-Driven Text-to-Image Generation (ICLR 2024 · similarity 0.87)
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.