Unlocking Few-Step Diffusion for Faithful Previews

Jing Jia; Sifan Liu; Guanyang Wang

cs.LGSep 28, 2026

Unlocking Few-Step Diffusion for Faithful Previews

Authors: Jing Jia, Sifan Liu, Guanyang Wang

Organizations: Department of Computer Science, Rutgers University · Department of Statistical Science, Duke University · Department of Statistics, Rutgers University

Abstract

Sampling latency compounds in diffusion workflows, where users generate and discard many candidates before keeping one. Surprisingly, the poor outputs of standard few-step samplers do not reflect a lack of reconstruction capacity: by optimizing only the initial noise, frozen 3-4-step samplers can closely reproduce their corresponding full-step outputs. Building on this finding, we learn corrections to the initial noise and denoising updates using endpoint supervision, improving correspondence with full-step outputs generated from the same noise and prompt. The resulting previews allow users to screen candidates cheaply and reserve full-step generation for promising ones. Input correction also transfers across sampling budgets without retraining. Experiments show substantial improvements in reference fidelity, including 53-78% lower reconstruction MSE than retrained LD3 on unconditional benchmarks, alongside improved ranking preservation and candidate selection on SD1.5, SDXL, and FLUX.1-dev.

Figures & tables

Figure 1: Traditional vs. preview-then-refinement workflow. Traditional generation requires full denoising steps (e.g. 60 steps) before selection, making rejected candidates expensive. Our enhancement module enables a faithful 3-step preview for cheap selection, with full generation performed only for selected candidates.

Figure 2: Input correction (A) and trajectory correction (B) for fast and faithful diffusion previews.

Figure 3: Noise optimization reveals the reconstruction capacity of frozen few-step samplers. Columns show raw outputs, full-step targets, and noise-optimized outputs.

Model / data	$N$	few/ref	Raw	Noise-Optimized	Gain (dB)
DDPM / CIFAR-10	100	3 / 1000	$17.00\pm 1.91$	$37.41\pm 1.89$	+20.41
DDPM / LSUN-Church	100	3 / 20	$14.96\pm 1.49$	$38.01\pm 3.15$	+23.05
DDPM / CelebA-HQ	100	3 / 20	$14.62\pm 1.65$	$39.14\pm 3.89$	+24.52
DDPM / LSUN-Bedroom	100	3 / 20	$15.52\pm 1.77$	$39.98\pm 3.09$	+24.46
FLUX.1-dev	4	4 / 28	$14.54\pm 1.95$	$37.49\pm 7.69$	+22.95

Table 1: Frozen-model oracle noise optimization. PSNR (dB) is the mean

\pm

sample SD over

N

images; few/ref denotes student/teacher steps.

Figure 4: Representative examples of input correction on CIFAR-10, LSUN-Church, and CelebA-HQ. Columns show uncorrected 3-step DDIM, 4-step DPM-Solver++, 4-step LD3, and our method.

Figure 5: PSNR and HPS on LSUN-Church and CelebA-HQ as the number of DDIM steps increases. Dashed lines are raw DDIMs.

Figure 6: Bedroom-to-CelebA-HQ correction.

Input Correction MSE ( $\times 10^{-3}$ ) $\downarrow$ / PSNR $\uparrow$ / LPIPS $\downarrow$ / SSIM $\uparrow$
Method	NFE	CIFAR-10				LSUN-Church				CelebA-HQ
3-step DDIM	3	20.79	17.281	0.355	0.575	33.48	15.010	0.571	0.580	34.64	14.882	0.583	0.666
4-step DDIM	4	15.36	18.652	0.296	0.654	24.11	16.447	0.482	0.632	24.81	16.388	0.417	0.729
4-step DPM++	4	10.21	20.655	0.279	0.717	7.77	21.445	0.341	0.742	8.14	21.563	0.301	0.812
4-step LD3 (transferred)	4	24.58	17.494	0.274	0.732	14.16	18.708	0.500	0.712	14.50	18.731	0.409	0.736
4-step LD3 (retrained)	4	9.75	21.045	0.248	0.764	6.23	22.487	0.343	0.762	5.08	23.847	0.215	0.864

Table 2: Within each dataset/CFG group, metrics are reported in the order MSE

\downarrow

, PSNR

\uparrow

, LPIPS

\downarrow

, and SSIM

\uparrow

. Noise-correction results are averaged over 100 paired test noises per dataset. For input correction,

3+1

denotes three denoiser evaluations and one lighter corrector evaluation. Trajectory correction is evaluated on SDXL and SD1.5. Bold denotes the best reported result within each group.

		SDXL Selected $\uparrow$ / Regret $\downarrow$ / Spearman $\uparrow$
Metric	Method	CFG = 1			CFG = 3			CFG = 5			CFG = 7
PickScore	3-step DDIM	19.642	1.377	0.016	21.724	0.915	0.009	22.162	0.855	-0.003	22.312	0.814	-0.011
	3-step DPM++	20.385	0.634	0.403	22.152	0.486	0.351	22.435	0.582	0.268	22.423	0.703	0.162
	3-step LD3	20.516	0.505	0.476	22.108	0.529	0.368	22.515	0.506	0.280	22.597	0.531	0.223
	Ours	20.621	0.397	0.593	22.215	0.424	0.463	22.658	0.359	0.466	22.741	0.386	0.347
ImageReward	3-step DDIM	-1.200	1.242	-0.030	0.274	0.701	-0.022	0.550	0.598	0.003	0.646	0.529	0.021

Table 3: Candidate-ranking preservation on SDXL and SD1.5 across CFG scales. We compare DDIM, DPM++, LD3, and our method, in addition ConSolver for SD1.5. For PickScore, ImageReward, and agent, we report the selected teacher score

\uparrow

, regret

\downarrow

, and Spearman rank correlation

\uparrow

Method	ImageReward $\uparrow$	PickScore $\uparrow$	HPS v2.1 $\uparrow$	CLIP $\uparrow$	AES $\uparrow$	NFE $\downarrow$
Random	0.9483	22.9300	29.6458	25.7252	5.6658	28
Diffusion Probe	0.9731	22.9311	29.7302	25.7060	5.6749	70
Probe-Select	0.9442	22.9432	29.6611	25.7982	5.6709	70
ConSolver	1.0085	22.9818	29.8980	25.7976	5.6857	60
TDD-FLUX	1.0331	23.0183	29.9616	25.9744	5.7166	60
Ours	1.1417	23.1869	30.3263	26.4693	5.7625	60

Table 4: FLUX.1-dev selection. HPS v2.1 and CLIP-L/14 cosine:

\times 100

. Bold: best non-oracle mean. NFE counts screening and final-generation denoiser calls, excluding decoding/scoring.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Metric order: MSE $\downarrow$ PSNR $\uparrow$ LPIPS $\downarrow$ SSIM $\uparrow$ HPS $\uparrow$ IR $\uparrow$
$K$	Method	LSUN-Church						CelebA-HQ
3	Raw DDIM	0.0336	15.024	0.571	0.577	13.010	-1.985	0.0355	14.738	0.592	0.658	15.958	-1.250
	Ours (all $p$ )	0.0031	25.661	0.287	0.850	18.624	-0.976	0.0015	28.909	0.183	0.915	23.482	-0.096
4	Raw DDIM	0.0245	16.386	0.489	0.626	15.859	-1.330	0.0254	16.242	0.425	0.720	19.976	-0.488
	Ours ( $p=1.7$ )	0.0039	24.486	0.244	0.864	19.081	-0.770	0.0026	26.318	0.160	0.904	24.499	-0.135
	Ours ( $p=2.0$ )	0.0031	25.562	0.235	0.872	18.957	-0.782	0.0017	28.309	0.146	0.922	24.333	-0.142

Appendix

Table 5: Transfer across steps on 200 samples per dataset. HPS: HPSv2.1

\times 100

; IR: ImageReward. Best values per dataset,

K

, and metric are bold.

CIFAR-10
Method	Calls	Params (M)	MSE $\downarrow$	PSNR $\uparrow$	LPIPS $\downarrow$	SSIM $\uparrow$
Raw DDIM3	3	0	21.18	17.29	0.149	0.578
Input correction	$3+1$	1.266	2.86	26.60	0.025	0.901
Trajectory correction	3	3.788	1.66	29.79	0.011	0.944
SD1.5, CFG 1
Raw DDIM3	3	0	20.05	17.33	0.801	0.402

Appendix

Table 6: Controlled comparison under matched data, endpoint supervision, and training-update budgets. MSE is multiplied by

10^{3}

; PSNR is in dB. Params counts trainable parameters, and Calls counts U-Net evaluations. The input arm uses one additional full U-Net corrector. References use 1,000 steps for CIFAR-10 and 100 steps for SD1.5. Results use 1,000 CIFAR-10 test inputs or 64 SD1.5 prompts with four noises each.

(a) Preference scores
Method	NFEs	PickScore $\uparrow$	HPS $\uparrow$	IR $\uparrow$
Raw DDIM3	$3$	$19.03\pm 0.45$	$16.53\pm 2.22$	$-0.98\pm 0.69$
Bedroom-backbone corrector + DDIM3	$3+1$	$\mathbf{20.37}\pm 0.54$	$\mathbf{27.34}\pm 1.77$	$\mathbf{0.54}\pm 0.36$
CelebA corrector + DDIM3	$3+1$	$19.90\pm 0.62$	$24.04\pm 2.13$	$0.10\pm 0.50$
DDIM20 teacher	$20$	$19.89\pm 0.61$	$24.88\pm 2.13$	$0.10\pm 0.51$
(b) Agreement with the DDIM20 teacher

Appendix

Table 7: Cross-backbone transfer from Bedroom to CelebA-HQ. Results are mean

\pm

SD over 200 samples. NFEs denote denoiser plus corrector forward passes. HPS denotes HPSv2.1

\times 100

; IR denotes ImageReward. Preference scores use the fixed text “a portrait photo of a person” for all unconditional outputs.

Figure 7: Comparison of FLUX.1-schnell and our method as previews of FLUX.1-dev. Columns show raw four-step Euler, Schnell, our corrected four-step sampler, and the 28-step reference. Although Schnell achieves higher ImageReward on the previews themselves, our method provides greater fidelity to the reference and better candidate-selection performance.

Figure 8: Noise norm normalization ablation. Columns show raw three-step DDIM, correction without normalization, complete correction, and the 20-step reference.

Figure 9: Correction direction ablation. Columns show raw three-step DDIM, reversed correction, another sample's correction direction rescaled to the original residual norm, complete correction, and the 20-step reference.

Figure 10: Attenuation when transferring a three-step corrector to eight-step DDIM. Columns show raw eight-step DDIM, correction with $\alpha=1$ , correction with $\alpha=(3/8)^{2.5}$ , and the 20-step reference.

Figure 11: Visual comparison of 60-step Reference images and 3-step DDIM, DPM++, LD3, and Ours generations using SDXL at CFG 1.

Figure 12: Visual comparison of 60-step Reference images and 3-step DDIM, DPM++, LD3, and Ours generations using SDXL at CFG 3.

Figure 13: Visual comparison of 60-step Reference images and 3-step DDIM, DPM++, LD3, and Ours generations using SDXL at CFG 5.

Figure 14: Visual comparison of 60-step Reference images and 3-step DDIM, DPM++, LD3, ConSolver, and Ours generations using Stable Diffusion 1.5 at CFG 1.

Figure 15: Visual comparison of 60-step Reference images and 3-step DDIM, DPM++, LD3, ConSolver, and Ours generations using Stable Diffusion 1.5 at CFG 3.

Figure 16: Visual comparison of 60-step Reference images and 3-step DDIM, DPM++, LD3, ConSolver, and Ours generations using Stable Diffusion 1.5 at CFG 5.

Figure 17: Comparison of 8 seed for the prompt “A large glass window in a living room.”

Figure 18: Comparison of 8 seed for the prompt “Comfortable, modern living room overlooking a wooded area”.

Figure 19: Comparison of 8 seed selection for the prompt “A living room with windows looking out onto a forest.”

Figure 20: Additional oracle noise examples.

Figure 21: Selected examples of input-correction transfer across sampling steps. Top: examples remaining close to the teacher across step counts. Bottom: examples approaching the teacher as the step count increases. All outputs reuse the same frozen three-step corrector with a fixed exponent of 2.5 and noise norm normalization; each row shares the same original noise.

Figure 22: Additional cross-backbone correction examples on CelebA-HQ. Each triplet shows raw three-step DDIM, the paired 20-step teacher, and three-step DDIM with the transferred input corrector. The corrector was trained through a Bedroom backbone using CelebA teacher targets.

Explore similar work

PreviewDiff: Multimodal Critic-Guided Search over Diffusion Latents
Sep 28, 2026Vighnesh Subramaniam, Boris Katz, Brian Cheung +3Diffusion Sampling Video Generation
MIT CSAIL · Google Cloud AI Research · UCSF
Simple Diffusion Language Models Are More Effective Few-Step Generators Than Reported
Sep 27, 2026Hasan Amin, Ming Yin, Rajiv KhannaDiffusion Language Models Few-Step Generation
Purdue University
FeatFix: Reuse What You Verify through Local Exact-Feature Correction for Faster Cached Diffusion Inference
Jul 30, 2026Hanshuai Cui, Zhiqing Tang, Zhi Yao +3Drafter Correction
School of Artificial Intelligence, Beijing Normal University, Beijing 100875, China · Institute of Artificial Intelligence and Future Networks, Beijing Normal University, Zhuhai 519087, China

Sep 28, 2026, cs.CVJ/K move · Enter open · S save

PreviewDiff: Multimodal Critic-Guided Search over Diffusion Latents

Vighnesh Subramaniam, Boris Katz, Brian Cheung, Chun-Liang Li+2

MIT CSAIL · Google Cloud AI Research · UCSF

Diffusion Sampling Video Generation Multimodal Search Agents Continuation

Open paper AskPDF arXiv

Diffusion models can produce striking images and videos, but they still struggle with the compositional details that make a generation faithful to a prompt, such as object counts, attribute binding, spatial relations, and temporally grounded actions. A common way to improve prompt satisfaction is to spend more compute at test time through Best-of-N sampling, but final-sample selection is fixed. Best-of-N can only choose among completed outputs and cannot repair a promising trajectory before it fails. We introduce PreviewDiff, a training-free test-time search method that turns diffusion sampling from scalar search into a multimodal critic-guided search over intermediate latents. At selected denoising checkpoints, PreviewDiff decodes a partial preview, asks a multimodal judge to score and critique it, and uses the resulting natural-language feedback to branch over semantic prompt edits and locally re-noised latent continuations. These branches are then scored and selectively rolled forward, allowing verifier compute to guide generation while the sample is still editable. Across image and video generation benchmarks, PreviewDiff consistently improves over budget-matched Best-of-N selection and strong scalar-search baselines. Ablations show that earlier interventions and increased search width provide the largest gains, while deeper search and additional semantic variants offer complementary improvements. PreviewDiff demonstrates that multimodal feedback is most useful not only as a final verifier, but as an active controller inside the denoising process.

Abstract

Explore similar work

PreviewDiff: Multimodal Critic-Guided Search over Diffusion Latents

Simple Diffusion Language Models Are More Effective Few-Step Generators Than Reported

FeatFix: Reuse What You Verify through Local Exact-Feature Correction for Faster Cached Diffusion Inference