Research map / Image generation / Text-to-image diffusion
Reusing Computation in Text-to-Image Diffusion for Efficient Generation of Image Sets
Dale Decatur, Thibault Groueix, Wang Yifan, Rana Hanocka, Vladimir Kim, Matheus Gadelha
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 |
|---|---|---|---|
| share early diffusion steps across clustered prompts at fixed quality | Kandinsky 2.2 · GenAI Bench | compute saved with similar VQA quality: 50.43% saved, 50.75% win | Improved |
| share early diffusion steps across clustered prompts at fixed quality | Kandinsky 2.2 · Style Variations | compute saved with similar VQA quality: 73.93% saved, 52.00% win | Improved |
| use semantic clustering of prompt embeddings instead of random clustering (ablation) | hierarchical diffusion · GenAI Bench | compute saved: 5.43% -> 50.43% | Improved |
| use a model trained with text-to-image prior (Kandinsky) instead of SD 1.5 (ablation) | hierarchical diffusion · Prompt Template | compute saved: 28.25% -> 69.25% | Improved |
| apply method to FLUX instead of Kandinsky (ablation) | hierarchical diffusion · Prompt Template | compute saved: 69.25% -> 3.61% | Hurt |
| share computation under fixed budget instead of standard diffusion | Kandinsky 2.2 · Prompt Template VQA Score | VQA Score at fixed compute budget | Improved |
| lower tau for earlier specialization (ablation) | hierarchical diffusion · Prompt Template | compute savings vs quality | Mixed |
| use non-linear (log) mapping for phi instead of linear (ablation) | hierarchical diffusion · text-to-image prompt sets | savings vs quality | Hurt |
| share computation on small homogeneous Style Variations set | hierarchical diffusion · Style Variations | CLIP diversity (lower is more diverse): 0.8071 -> 0.8374 | Hurt |
| share computation on large prompt sets (500+) | hierarchical diffusion · GenAI Bench / Prompt Template / Animals | CLIP diversity: GenAI Bench 0.6082 -> 0.6027 | No clear effect |
Closest papers by meaning
- Würstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models (ICLR 2024 · similarity 0.92)
- On Discrete Prompt Optimization for Diffusion Models (ICML 2024 · similarity 0.92)
- Prompting Hard or Hardly Prompting: Prompt Inversion for Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.91)
- Prompt-Free Diffusion: Taking "Text" out of Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.91)
- Not All Prompts Are Made Equal: Prompt-based Pruning of Text-to-Image Diffusion Models (ICLR 2025 · similarity 0.91)
- On improved Conditioning Mechanisms and Pre-training Strategies for Diffusion Models (NeurIPS 2024 · similarity 0.91)
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.
