AI Paper Atlas

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

ICCV 2023 · Image generation › Text-to-image diffusion · approach: Text-driven synthesis and semantic guidance · 64 citations (OpenAlex)

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 changedWhereMeasuredEffect
optimize diffusion latents through the full EDICT chain instead of one-step approximated classifier guidanceStable Diffusion with CLIP guidance · DrawBench DALLECLIP score: +3.1% relativeImproved
optimize diffusion latents through full chain instead of one-step guidanceStable Diffusion with CLIP guidance · DrawBench RedditCLIP score: +2.6% relativeImproved
use DOODL instead of one-step guidance for human-judged prompt alignmentStable Diffusion · DrawBench Reddit human evaluationPASR: +11.3% over vanilla SD, +6% over one-step guidanceImproved
use DOODL with fine-grained classifiers instead of one-step guidance for vocabulary expansionStable Diffusion · CUB, Dogs, Aircraftper-class FID change (%): -3.2% vs +6.1% on CUBImproved
use DOODL for fine-grained aircraft classesStable Diffusion · FGVC-Aircraftper-class FID change (%): -0.4%No clear effect
use DOODL for on-the-fly personalization with CLIP image embedding lossStable Diffusion · dog personalization human studysuccess rate (majority): 1% -> 14% over baseline guidanceImproved
use DOODL with aesthetic scorer for text-to-image generation versus original generationsStable Diffusion · 8 prompts x 16 seeds human studywin rate: win-draw-loss 0.58-0.11-0.31Improved
use DOODL instead of one-step classifier guidance for aesthetic scoringStable Diffusion · aesthetic text-to-image human studywin rate: win-draw-loss 0.48-0.07-0.45No clear effect
lower similarity bar for personalization judgementDOODL personalization · dog personalization human studysuccess rate (aggregate): DOODL 52.1% vs baseline 28.5% vs original 12.2%Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy