Research map / Image generation / Text-to-image diffusion
End-to-End Diffusion Latent Optimization Improves Classifier Guidance
Bram Wallace, Akash Gokul, Stefano Ermon, Nikhil Naik
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 |
|---|---|---|---|
| optimize diffusion latents through the full EDICT chain instead of one-step approximated classifier guidance | Stable Diffusion with CLIP guidance · DrawBench DALLE | CLIP score: +3.1% relative | Improved |
| optimize diffusion latents through full chain instead of one-step guidance | Stable Diffusion with CLIP guidance · DrawBench Reddit | CLIP score: +2.6% relative | Improved |
| use DOODL instead of one-step guidance for human-judged prompt alignment | Stable Diffusion · DrawBench Reddit human evaluation | PASR: +11.3% over vanilla SD, +6% over one-step guidance | Improved |
| use DOODL with fine-grained classifiers instead of one-step guidance for vocabulary expansion | Stable Diffusion · CUB, Dogs, Aircraft | per-class FID change (%): -3.2% vs +6.1% on CUB | Improved |
| use DOODL for fine-grained aircraft classes | Stable Diffusion · FGVC-Aircraft | per-class FID change (%): -0.4% | No clear effect |
| use DOODL for on-the-fly personalization with CLIP image embedding loss | Stable Diffusion · dog personalization human study | success rate (majority): 1% -> 14% over baseline guidance | Improved |
| use DOODL with aesthetic scorer for text-to-image generation versus original generations | Stable Diffusion · 8 prompts x 16 seeds human study | win rate: win-draw-loss 0.58-0.11-0.31 | Improved |
| use DOODL instead of one-step classifier guidance for aesthetic scoring | Stable Diffusion · aesthetic text-to-image human study | win rate: win-draw-loss 0.48-0.07-0.45 | No clear effect |
| lower similarity bar for personalization judgement | DOODL personalization · dog personalization human study | success rate (aggregate): DOODL 52.1% vs baseline 28.5% vs original 12.2% | Improved |
Closest papers by meaning
- Elucidating the design space of classifier-guided diffusion generation (ICLR 2024 · similarity 0.91)
- Plug-and-Play Diffusion Distillation (CVPR 2024 · similarity 0.89)
- Unlocking Dataset Distillation with Diffusion Models (NeurIPS 2025 · similarity 0.87)
- Adaptive Guidance: Training-free Acceleration of Conditional Diffusion Models (AAAI 2025 · similarity 0.87)
- Universal Guidance for Diffusion Models (ICLR 2024 · similarity 0.87)
- Angle Domain Guidance: Latent Diffusion Requires Rotation Rather Than Extrapolation (ICML 2025 · 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.
