FASTER: Fast Adjoint Stochastic Transport for Endpoint Refinement in Reward-Guided Image Editing

Yimiao Zhou; Zejia Zhong; Jingya Wang; Ye Shi

cs.CVOct 3, 2026

FASTER: Fast Adjoint Stochastic Transport for Endpoint Refinement in Reward-Guided Image Editing

Authors: Yimiao Zhou, Zejia Zhong, Jingya Wang, Ye Shi

Organizations: ShanghaiTech University Shanghai, China

Abstract

Reward-guided image editing at test time seeks to improve a specified reward while preserving source content and visual plausibility. Many existing approaches optimize candidates through pretrained generation processes, making repeated adjustment depend on costly large-model execution and, in some cases, backbone backpropagation. We develop a theoretical framework that jointly accounts for reward, source preservation, and pretrained-prior preferences, allowing the desired output distribution to be specified separately from the dynamics used to realize it. Based on this framework, we introduce FASTER, which trains a small network for each source and objective to perform inexpensive editing, while pretrained and reward models provide feedback on candidate outputs. By reusing each candidate and its feedback across multiple small-network updates, FASTER reduces repeated sampling and supervision queries without placing the pretrained generative backbone inside the inner optimization loop. On SD3, FASTER leads all four target metrics and several validation metrics among the evaluated methods. Compared with the evaluated baseline that optimizes controls along pretrained generation trajectories, FASTER achieves editing-time speedups of up to ${6.91\times}$ on Stable Diffusion 3 and ${24.14\times}$ on Stable Diffusion 1.5.

Figures & tables

Figure 1: Decoupling pretrained knowledge from editing dynamics. (a) Trajectory optimization repeatedly executes a frozen generative backbone. (b) FASTER learns a small per-image editor at test time, reusing frozen-model feedback across updates; the learned network then performs the edit.

Figure 2: FASTER training and editing pipeline. (a) Each of $K$ refreshes caches an endpoint and its detached preservation, prior, and reward target for $M$ controller updates. (b) After training, a fresh controller rollout is decoded into the edited image.

Human preference
Metric	ITOC	GA	DPS	FreeDoM	TFG	MPGD	FASTER
TARGET ImageReward $\uparrow$	1.8326	1.8645	1.4687	1.6535	1.5128	0.9413	1.8958
[5pt/3pt] VALIDATION HPSv2 $\uparrow$	0.2619	0.2608	0.2579	0.2576	0.2580	0.2563	0.2624
VALIDATION CLIPScore $\uparrow$	0.2734	0.2776	0.2566	0.2388	0.2539	0.2655	0.2783
VALIDATION Aesthetic $\uparrow$	5.7933	5.6329	5.5978	5.6719	5.4951	5.4667	5.6991
Style transfer

Table 1: Results on Stable Diffusion 3. Each task includes a target metric, which is directly optimized, and validation metrics used to assess potential reward hacking. Arrows indicate whether higher (

\uparrow

) or lower (

\downarrow

) values are better. Bold values indicate the best result for each metric.

Figure 3: Editing time of ITOC and FASTER across four tasks on SD3 and SD1.5.

Figure 4: Qualitative comparisons on four SD3 editing tasks. Rows show text-guided image editing, counterfactual generation, human preference enhancement, and style transfer. For text-guided image editing, the target prompt is supplied to the pretrained generative model of each applicable method. The style reference is shown on the left.

Figure 5: FASTER with CLIPScore optimization or the Editing Term for the same sunglasses request. Values are CLIPScore.

Figure 6: Effect of

K

and four parameter settings on the SD3 ImageReward task. Each curve averages the same three source–prompt pairs at saved checkpoints from one run per setting. All settings use

\alpha\tau^{2}=1

and

c_{p}=0

. Solid curves use

\tau=0.0125

; the dashed curve uses

\tau=0.025

. The dotted vertical line marks

K=20

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Backbone	Task	$\alpha\tau^{2}$	$\tau$	$c_{r}$	$c_{p}$
SD1.5	Human	1	0.0125	320	0.1
SD1.5	Style	0.01	0.002	10240	0
SD1.5	Counterfactual	1	0.002	2560	0
SD1.5	Facial edit	1	0.00625	1280	0.01
SD3	Human	1	0.0125	320	0
SD3	Style	0.01	0.0125	5120	0

Appendix

Table 2: Selected FASTER settings for the quantitative benchmarks.

Model	Task	ITOC $(n,w,\ell)$	GA $\ell$	DPS $\rho$	FreeDoM $\rho$	TFG $(\rho,\mu)$	MPGD $\lambda$
SD3	Human	$(15,500,0.005)$	2	5	5	$(5,1)$	1
SD3	Style	$(15,1000,0.005)$	3	50	20	$(40,2.5)$	12
SD3	Counterfactual	$(15,200,0.005)$	1	7	3	$(7,0.5)$	7
SD3	Face	$(15,1000,0.005)$	1.5	75	30	$(60,2.5)$	10
SD1.5	Human	$(20,500,0.005)$	2	3	1	$(1,0.5)$	1
SD1.5	Style	$(20,200,0.01)$	3	15	7.5	$(10,1)$	12

Appendix

Table 3: Task-specific baseline settings used for the reported results. The notation and shared settings are defined in the text.

Human preference
Metric	ITOC	GA	DPS	FreeDoM	TFG	MPGD	FASTER
TARGET ImageReward $\uparrow$	1.8981	1.9532	1.5374	1.3118	1.6908	1.1938	1.9316
[5pt/3pt] VALIDATION HPSv2 $\uparrow$	0.2666	0.2615	0.2578	0.2555	0.2606	0.2535	0.2649
VALIDATION CLIPScore $\uparrow$	0.2773	0.2743	0.2546	0.2366	0.2642	0.2531	0.2798
VALIDATION Aesthetic $\uparrow$	5.8058	5.5443	5.4939	5.4869	5.5269	5.2346	5.8075
Style transfer

Appendix

Table 4: Results on Stable Diffusion 1.5. Each task includes a target metric, which is directly optimized, and validation metrics used to assess potential reward hacking. Arrows indicate whether higher (

\uparrow

) or lower (

\downarrow

) values are better. Bold values indicate the best result for each metric.

Figure 7: Editing time on one H20 across four tasks and two model families, averaged over three images per setting. The data and timing scope match Fig. 3 ; the vertical axis is logarithmic.

Figure 8: Additional matched counterfactual edits. Source and target classes are shown at left; method columns follow Fig. 4 .

Figure 9: Additional matched facial edits. The requested text prompts are shown at left; method columns follow Fig. 4 .

Figure 10: Additional matched human-preference examples, with the source image and outputs from each method.

Figure 11: Additional style-transfer examples. The style reference and source image occupy the first two columns; method columns follow Fig. 4 .

Figure 12: Adding sunglasses to the same puppy with the Editing Term. Columns vary the noise-time interval and rows vary $c_{e}$ ; the source is repeated for comparison.

Figure 13: Saved outer-loop checkpoints for the puppy sunglasses edit at $c_{e}=250$ and $t\in[0.75,0.90]$ .

Panel of Figure 14 — Figure 14: Editing Term strength and noise-time sweeps for lipstick (left) and apparent age (right). The source is repeated at left of each row.

Figure 15: Endpoint feedback reuse across four SD3 tasks. Curves show three-image means as the number of outer refreshes

K

increases; colors denote the number of controller updates

M

per refresh. The first panel in each row is the optimized target, and the remaining panels are validation metrics.

Image Editing

Image editing changes selected visual content while preserving the parts of an image that should remain unchanged.

See all 318 papers

Inverse-and-Edit: Simple and Effective Framework for Fast Image Editing
Jun 23, 2025Ilia Beletskii, Andrey Kuznetsov, Aibek AlanovText-Guided Image Editing Diffusion-Based Image Editing
HSE University, AIRI · AIRI, Sber, Innopolis
Refinement Is Inherently Editable: Training-Free Prompt-to-Prompt Image Editing with Generative Refinement Network
Sep 17, 2026Yulong Chen, Ziqian Zhang, Haoyu Zhang +4Text-Guided Image Editing Training-Free Image Editing
City University of Hong Kong (Dongguan), Guangdong, China · City University of Hong Kong, Hong Kong, China · Mohamed bin Zayed University of Artificial Intelligence, Masdar, Abu Dhabi
RewardFlow: Generate Images by Optimizing What You Reward
Apr 9, 2026Onkar Susladkar, Dong-Hwan Jang, Tushar Prakash +7Diffusion Model Guidance Diffusion-Based Image Editing
University of Illinois Urbana-Champaign · Sony Research, India

Jun 23, 2025J/K move · Enter open · S save

Inverse-and-Edit: Simple and Effective Framework for Fast Image Editing

Ilia Beletskii, Andrey Kuznetsov, Aibek Alanov

HSE University, AIRI · AIRI, Sber, Innopolis

Text-Guided Image Editing Diffusion-Based Image Editing Diffusion Model Adaptation

Open paper AskPDF arXiv

Diffusion-based text-to-image models have recently been distilled into few-step samplers and fast generators, reducing inference cost while maintaining high image quality. However, text-guided editing with such accelerated backbones remains challenging: inversion in a small number of steps is unreliable, local errors accumulate along the trajectory, and existing fast-editing methods typically rely on backbone-specific inversion procedures, which limits their transferability. In this work, we propose a simple and universal framework for text-guided image editing with arbitrary accelerated diffusion and flow-matching models. Instead of optimizing local, per-timestep objectives, we directly minimize a global editing objective over the entire inversion-generation trajectory. The framework fine-tunes only 1.5-7 % of the base-model parameters through two lightweight LoRA adapters (forward and backward), trained on fewer than 10 K triplets from the publicly available HQ-Edit dataset. We validate our method on three architecturally distinct backbones: consistency-distilled SD1.5 LCM-Dreamshaper, adversarial-distilled SDXL-Turbo, and rectified-flow-distilled Z-Image-Turbo. On PIE-Bench, our method achieves the strongest editing alignment within each backbone family while ensuring a strong level of content preservation. Inference adds no extra steps beyond the standard inversion and editing passes. The project page with visual results is available at iliabeletskii.github.io/inverse-and-edit.

Abstract

Image Editing

Explore similar work

Inverse-and-Edit: Simple and Effective Framework for Fast Image Editing

Refinement Is Inherently Editable: Training-Free Prompt-to-Prompt Image Editing with Generative Refinement Network

RewardFlow: Generate Images by Optimizing What You Reward