Matches in SemOpenAlex for { <https://semopenalex.org/work/W4387294554> ?p ?o ?g. }
- W4387294554 abstract "We present Direct Reward Fine-Tuning (DRaFT), a simple and effective method for fine-tuning diffusion models to maximize differentiable reward functions, such as scores from human preference models. We first show that it is possible to backpropagate the reward function gradient through the full sampling procedure, and that doing so achieves strong performance on a variety of rewards, outperforming reinforcement learning-based approaches. We then propose more efficient variants of DRaFT: DRaFT-K, which truncates backpropagation to only the last K steps of sampling, and DRaFT-LV, which obtains lower-variance gradient estimates for the case when K=1. We show that our methods work well for a variety of reward functions and can be used to substantially improve the aesthetic quality of images generated by Stable Diffusion 1.4. Finally, we draw connections between our approach and prior work, providing a unifying perspective on the design space of gradient-based fine-tuning algorithms." @default.
- W4387294554 created "2023-10-03" @default.
- W4387294554 creator A5008480143 @default.
- W4387294554 creator A5026208886 @default.
- W4387294554 creator A5027477681 @default.
- W4387294554 creator A5035123135 @default.
- W4387294554 date "2023-09-29" @default.
- W4387294554 modified "2023-10-04" @default.
- W4387294554 title "Directly Fine-Tuning Diffusion Models on Differentiable Rewards" @default.
- W4387294554 doi "https://doi.org/10.48550/arxiv.2309.17400" @default.
- W4387294554 hasPublicationYear "2023" @default.
- W4387294554 type Work @default.
- W4387294554 citedByCount "0" @default.
- W4387294554 crossrefType "posted-content" @default.
- W4387294554 hasAuthorship W4387294554A5008480143 @default.
- W4387294554 hasAuthorship W4387294554A5026208886 @default.
- W4387294554 hasAuthorship W4387294554A5027477681 @default.
- W4387294554 hasAuthorship W4387294554A5035123135 @default.
- W4387294554 hasBestOaLocation W43872945541 @default.
- W4387294554 hasConcept C105795698 @default.
- W4387294554 hasConcept C106131492 @default.
- W4387294554 hasConcept C111472728 @default.
- W4387294554 hasConcept C115961682 @default.
- W4387294554 hasConcept C121332964 @default.
- W4387294554 hasConcept C121955636 @default.
- W4387294554 hasConcept C126255220 @default.
- W4387294554 hasConcept C12713177 @default.
- W4387294554 hasConcept C134306372 @default.
- W4387294554 hasConcept C136197465 @default.
- W4387294554 hasConcept C138885662 @default.
- W4387294554 hasConcept C14036430 @default.
- W4387294554 hasConcept C140779682 @default.
- W4387294554 hasConcept C144133560 @default.
- W4387294554 hasConcept C153258448 @default.
- W4387294554 hasConcept C154945302 @default.
- W4387294554 hasConcept C19499675 @default.
- W4387294554 hasConcept C196083921 @default.
- W4387294554 hasConcept C202615002 @default.
- W4387294554 hasConcept C2779530757 @default.
- W4387294554 hasConcept C2780586882 @default.
- W4387294554 hasConcept C2781249084 @default.
- W4387294554 hasConcept C31972630 @default.
- W4387294554 hasConcept C33923547 @default.
- W4387294554 hasConcept C41008148 @default.
- W4387294554 hasConcept C50644808 @default.
- W4387294554 hasConcept C52740198 @default.
- W4387294554 hasConcept C69357855 @default.
- W4387294554 hasConcept C78458016 @default.
- W4387294554 hasConcept C86803240 @default.
- W4387294554 hasConcept C97355855 @default.
- W4387294554 hasConcept C97541855 @default.
- W4387294554 hasConcept C99498987 @default.
- W4387294554 hasConceptScore W4387294554C105795698 @default.
- W4387294554 hasConceptScore W4387294554C106131492 @default.
- W4387294554 hasConceptScore W4387294554C111472728 @default.
- W4387294554 hasConceptScore W4387294554C115961682 @default.
- W4387294554 hasConceptScore W4387294554C121332964 @default.
- W4387294554 hasConceptScore W4387294554C121955636 @default.
- W4387294554 hasConceptScore W4387294554C126255220 @default.
- W4387294554 hasConceptScore W4387294554C12713177 @default.
- W4387294554 hasConceptScore W4387294554C134306372 @default.
- W4387294554 hasConceptScore W4387294554C136197465 @default.
- W4387294554 hasConceptScore W4387294554C138885662 @default.
- W4387294554 hasConceptScore W4387294554C14036430 @default.
- W4387294554 hasConceptScore W4387294554C140779682 @default.
- W4387294554 hasConceptScore W4387294554C144133560 @default.
- W4387294554 hasConceptScore W4387294554C153258448 @default.
- W4387294554 hasConceptScore W4387294554C154945302 @default.
- W4387294554 hasConceptScore W4387294554C19499675 @default.
- W4387294554 hasConceptScore W4387294554C196083921 @default.
- W4387294554 hasConceptScore W4387294554C202615002 @default.
- W4387294554 hasConceptScore W4387294554C2779530757 @default.
- W4387294554 hasConceptScore W4387294554C2780586882 @default.
- W4387294554 hasConceptScore W4387294554C2781249084 @default.
- W4387294554 hasConceptScore W4387294554C31972630 @default.
- W4387294554 hasConceptScore W4387294554C33923547 @default.
- W4387294554 hasConceptScore W4387294554C41008148 @default.
- W4387294554 hasConceptScore W4387294554C50644808 @default.
- W4387294554 hasConceptScore W4387294554C52740198 @default.
- W4387294554 hasConceptScore W4387294554C69357855 @default.
- W4387294554 hasConceptScore W4387294554C78458016 @default.
- W4387294554 hasConceptScore W4387294554C86803240 @default.
- W4387294554 hasConceptScore W4387294554C97355855 @default.
- W4387294554 hasConceptScore W4387294554C97541855 @default.
- W4387294554 hasConceptScore W4387294554C99498987 @default.
- W4387294554 hasLocation W43872945541 @default.
- W4387294554 hasOpenAccess W4387294554 @default.
- W4387294554 hasPrimaryLocation W43872945541 @default.
- W4387294554 hasRelatedWork W1965320053 @default.
- W4387294554 hasRelatedWork W2002748013 @default.
- W4387294554 hasRelatedWork W2006722592 @default.
- W4387294554 hasRelatedWork W3171174687 @default.
- W4387294554 hasRelatedWork W3202114911 @default.
- W4387294554 hasRelatedWork W4221151455 @default.
- W4387294554 hasRelatedWork W4285069894 @default.
- W4387294554 hasRelatedWork W4307415184 @default.
- W4387294554 hasRelatedWork W4372259970 @default.
- W4387294554 hasRelatedWork W4382498635 @default.
- W4387294554 isParatext "false" @default.
- W4387294554 isRetracted "false" @default.