ReAL: Accelerating Flow Matching through Segment Advancement with Shared Lookahead

Xuanhua Yin; Chuanzhi Xu; Haoxian Zhou; Shunqi Mao; Weidong Cai

cs.CVSep 27, 2026

ReAL: Accelerating Flow Matching through Segment Advancement with Shared Lookahead

Authors: Xuanhua Yin, Chuanzhi Xu, Haoxian Zhou, Shunqi Mao, Weidong Cai

Organizations: School of Computer Science, The University of Sydney

Abstract

Flow-matching models generate high-quality images and videos, but repeated neural network evaluations make sampling expensive. Skipping evaluations reduces this cost by extending an available velocity estimate over a longer span. However, local velocity agreement alone does not determine a suitable span, and checking each candidate endpoint adds costly model calls. We introduce ReAL, a training-free sampler that selects how far to advance using one shared lookahead. Our key insight is that the discrepancy between uncorrected and lookahead-corrected endpoint proposals can be computed directly from the observed velocity mismatch and the candidate span beyond the lookahead. This relation provides a span-dependent selection criterion without additional endpoint evaluations. The same lookahead selects the longest passing candidate span, corrects the accepted update, and supplies its velocity as the next starting estimate. After initialization, each regular iteration requires only one fresh evaluation. ReAL uses the pretrained velocity output and original noise schedule, with no additional training or access to internal features. Experiments cover four image-generation backbones, video generation, and image editing. ReAL achieves 4.91x measured speedup on FLUX.1-dev while retaining 97.0% of dense mean ImageReward. On HunyuanVideo, it achieves a 5.49x speedup while maintaining a VBench score close to that of dense sampling.

Figures & tables

Figure 1: One lookahead for selection, correction, and reuse. Left: a schematic of reference sampling, extrapolation, and the three roles of shared lookahead in ReAL. Right: FLUX.1-dev outputs with the same prompt and initial noise, using dense sampling (50 evaluations), fixed-stride sampling (13), and ReAL (9).

Figure 2: One lookahead serves three roles: segment selection, endpoint correction, and velocity reuse. The blue carry supplies stored $u_{i}$ , and the fresh query at $z_{i}$ produces shared $w_{i}$ . State images are illustrative. The regular iteration is shown here, with initialization and tail handling specified in Algorithm 1 .

FLUX.1-dev	Efficiency					Image Reward $\uparrow$	CLIP Score $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
FLUX.1-dev	Latency (s) $\downarrow$	Speedup $\uparrow$	FLOPs ( $10^{12}$ ) $\downarrow$	FLOPs ratio $\uparrow$	Mem (GB) $\downarrow$	Image Reward $\uparrow$	CLIP Score $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
50 steps	11.38	1.00 $\times$	2916.7	1.00 $\times$	31.92	0.9907	18.73	-	-	-
40% steps	4.50	2.53 $\times$	1126.1	2.59 $\times$	31.92	0.9468	18.31	16.88	0.7378	0.3216
30% steps	3.49	3.26 $\times$	875.89	3.33 $\times$	31.92	0.9381	17.92	15.03	0.6523	0.4279
22% steps	2.60	4.38 $\times$	655.4	4.45 $\times$	31.92	0.8917	17.36	13.79	0.6032	0.4683
TeaCache 1	3.91	2.91 $\times$	982.1	2.97 $\times$	31.97	0.9724	17.85	16.09	0.6679	0.3964

Table 1: FLUX.1-dev text-to-image comparison. Results use 200 DrawBench prompts, five seeds, and

1024^{2}

resolution. Superscripts

1

2

, and

3

denote progressively faster operating points. Bold marks the best accelerated quality value within each tier. Dashes denote unavailable dense-reference similarity fields.

Method	Efficiency						Generation quality
Method	Latency (s) $\downarrow$	Speedup $\uparrow$	FLOPs ( $10^{12}$ ) $\downarrow$	FLOPs ratio $\uparrow$	NFE $\downarrow$	Mem (GB) $\downarrow$	VBench $\uparrow$	PSNR $\uparrow$	SSIM $\uparrow$	LPIPS $\downarrow$
Dense 50 steps	298.41	1.00 $\times$	23750.0	1.00	50.0	60.44	83.27	N/A	N/A	N/A
22% steps	63.47	4.70 $\times$	4979.0	4.77	10.7	60.44	80.36	19.49	0.7002	0.4228
TeaCache †	84.12	3.55 $\times$	5549.1	4.28	11.5	60.56	81.39	21.94	0.7339	0.2077
TaylorSeer †	79.73	3.74 $\times$	5031.8	4.72	10.7	93.09	82.73	20.21	0.7549	0.2557
DiCache †	98.81	3.02 $\times$	6805.2	3.49	14.1	60.48	82.87	25.96	0.8206	0.2020

Table 2: HunyuanVideo on VBench. Results use 946 prompts, five seeds,

544\times 960

, and 65 frames. Symbols

\dagger

and

\ddagger

denote medium and high acceleration. Step reduction is the control. Bold marks the best accelerated result within each tier. Latency is median event time, speedup uses the matched dense run, and Mem is peak allocation. Appendix B gives the protocol.

Figure 3: FLUX.1-dev qualitative comparison. ReAL retains the requested three-object composition, cyan-haired character, and literal word “SEAL” in these matched examples. The rows use identical prompts, initial noise, and text embeddings across methods. Each method uses one fixed configuration. Appendix F gives the full prompts, settings, and additional comparisons.

Figure 4: HunyuanVideo qualitative comparison. Matched-seed bf16 runs compare ReAL with seven reference methods at frames 0, 32, and 64. Table 2 uses separate fp16 timing runs. Appendix G gives all fifteen methods and settings.

Figure 5: Editing quality across acceleration settings. Step1X-Edit on 606 examples with five shared seeds. Lines connect measured settings, and dashed lines mark dense quality. Table 5 reports all 22 settings.

Verification	Speedup $\uparrow$	IR $\uparrow$	NFE $\downarrow$
None	9.86 $\times$	$-0.0168$	5.0
Local only	7.96 $\times$	0.4395	6.0
Span only	4.58 $\times$	0.9621	9.8
Both criteria	4.20 $\times$	0.9864	10.3
Dense verification	0.42 $\times$	0.7801	118.4

Table 3: Selection, correction, and cross-segment reuse on FLUX. Panels (a) and (b) use a separate fallback-only diagnostic protocol and should not be read as replacements for the final current-method comparison in Table 1 . Panel (c) is a 100-prompt carry-policy comparison at

q_{\rm seg}=1.65

. The appendix reports full settings and statistics.

Figure 6: Quality versus compute on three additional backbones. On DrawBench-100, ReAL uses

K=6

and a threshold sweep per backbone. Points are measured settings, and lines guide the eye. The dense-head line connects its best measured trade-offs, with other settings shown faintly. TeaCache uses output-delta reuse on the SD3 backbones, and AdaptiveDiffusion uses the flow adaptation on SD3-Medium. Numerical comparisons appear in Appendix C .

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Video medium	Video high	Editing Tier 1	Editing Tier 2	Editing Tier 3
TeaCache	ell=.31	ell=.40	ell=.20	ell=.32	ell=.41
TaylorSeer	N=4	N=5	N=3	N=4	N=5
DiCache	t=.71	t=.80	t=.60	N/A	N/A
SpeCa	ts=3	ts=6	ts=2	ts=4	ts=6
DPCache	K=13	K=9	K=16	K=11	K=8
Spectrum	ws=2	ws=4	ws=2	ws=3	ws=5

Appendix

Table 4: Operating points. Video columns denote medium and high acceleration. Editing columns denote Tiers 1, 2, and 3. Baseline settings use method-specific text labels from the experiment records, and these labels have no shared mathematical meaning across rows. N/A indicates an unreported point.

Method	Efficiency				Editing quality
Method	Latency (s) $\downarrow$	Speedup $\uparrow$	NFE $\downarrow$	Mem (GB) $\downarrow$	$G_{\rm SC}\uparrow$	$G_{\rm PQ}\uparrow$	$G_{\rm O}\uparrow$
Dense	6.842	1.00 $\times$	28.00	12.04	5.24	5.31	4.20
TaylorSeer 1	2.321	2.95 $\times$	9.21	12.24	4.93	4.81	4.05
TeaCache 1	2.293	2.98 $\times$	9.40	12.75	4.89	4.97	3.96
SpeCa 1	2.720	2.52 $\times$	10.90	12.30	4.99	4.94	4.00
Spectrum 1	1.734	3.95 $\times$	6.80	12.26	4.83	5.06	4.00

Appendix

Table 5: Step1X-Edit on GEdit-Bench. Evaluation of the current ReAL controller: 606 examples, five seeds,

512^{2}

G_{\rm SC}

G_{\rm PQ}

, and

G_{\rm O}

denote semantic consistency, perceptual quality, and overall score. Superscripts

1

2

, and

3

denote acceleration tiers. Bold marks the best accelerated overall score within each numbered tier. FlowCast’s

\dagger

marks its non-accelerating setting. Latency is median event time, and Mem is peak allocated memory.

Method	Subj.	Bg.	Flicker	Smooth	Dynamic	Aesthetic	Imaging	Object
Dense 50 steps	95.36	97.76	99.44	98.99	71.11	60.36	70.10	86.10
22% steps	96.15	98.26	99.56	99.31	43.33	58.50	67.05	81.11
TeaCache	96.12	98.23	99.67	99.35	57.50	58.59	67.22	81.27
TaylorSeer	95.58	97.91	99.52	99.11	66.67	59.80	69.27	84.84
DiCache	95.58	97.87	99.51	99.08	67.22	60.02	69.73	84.98
SpeCa	95.61	97.95	99.54	99.13	64.17	59.82	69.22	84.72

Appendix

Table 6: Medium acceleration: VBench dimensions (a). Raw dimension scores

\times 100

. Dense and direct step reduction are shared references.

Method	Multiple	Action	Color	Spatial	Scene	App. style	Temp. style	Overall cons.
Dense 50 steps	71.19	94.20	88.70	68.68	53.88	19.80	23.89	26.44
22% steps	62.44	90.80	84.51	61.49	48.27	18.99	22.96	25.41
TeaCache	64.06	91.00	84.74	61.97	48.88	19.13	22.97	25.31
TaylorSeer	69.39	93.40	87.68	66.98	52.55	19.62	23.63	26.15
DiCache	69.52	93.60	87.87	67.51	52.76	19.66	23.70	26.25
SpeCa	68.86	93.20	87.31	66.69	51.56	19.58	23.60	26.08

Appendix

Table 7: Medium acceleration: VBench dimensions (b). Raw dimension scores

\times 100

. Dense and direct step reduction are shared references.

Method	Subj.	Bg.	Flicker	Smooth	Dynamic	Aesthetic	Imaging	Object
Dense 50 steps	95.36	97.76	99.44	98.99	71.11	60.36	70.10	86.10
22% steps	96.15	98.26	99.56	99.31	43.33	58.50	67.05	81.11
TeaCache	96.38	98.38	99.68	99.41	50.00	57.92	66.25	79.36
TaylorSeer	96.02	98.16	99.63	99.26	57.22	58.47	67.56	82.00
DiCache	96.19	98.20	99.62	99.27	55.83	58.84	67.54	81.15
SpeCa	96.13	98.21	99.63	99.30	48.33	58.64	67.07	81.16

Appendix

Table 8: High acceleration: VBench dimensions (a). Raw dimension scores

\times 100

. Dense and direct step reduction are shared references.

Method	Multiple	Action	Color	Spatial	Scene	App. style	Temp. style	Overall cons.
Dense 50 steps	71.19	94.20	88.70	68.68	53.88	19.80	23.89	26.44
22% steps	62.44	90.80	84.51	61.49	48.27	18.99	22.96	25.41
TeaCache	61.07	89.40	82.03	60.26	46.35	18.72	22.56	24.83
TaylorSeer	64.60	91.00	85.20	63.26	49.61	19.19	23.09	25.47
DiCache	63.56	91.20	85.24	63.26	49.25	19.20	23.06	25.52
SpeCa	63.61	90.40	84.17	62.22	47.07	19.04	22.89	25.36

Appendix

Table 9: High acceleration: VBench dimensions (b). Raw dimension scores

\times 100

. Dense and direct step reduction are shared references.

Figure 7: VBench dimension profiles at medium and high acceleration. Columns separate the original seven competing accelerators from the five additions. The same dense and ReAL references appear in both. Quality axes show raw dimension scores multiplied by 100 without per-axis rescaling. The speed axis uses

100\%=5.49\times

. All panels share axis order and limits. The complete 16-dimensional tables also report dynamic degree and imaging quality. Curves show the reported dimension scores. The export provides configuration-level aggregates rather than per-seed intervals.

	$q_{\rm seg}=2.6$ vs. rule $3{:}K$				$q_{\rm seg}=3.8$ vs. rule $2{:}K$
$K$	ReAL		Rule		ReAL		Rule
	NFE	IR	NFE	IR	NFE	IR	NFE	IR
12	7.65	0.884	8.00	0.893	6.32	0.676	7.00	0.831
16	7.76	0.853	7.00	0.700	6.16	0.635	6.00	0.578
24	7.83	0.904	6.00	0.283	6.41	0.738	5.00	0.034

Appendix

Table 10: Segment-cap study on FLUX. Absolute mean IR on 100 matched prompt and seed pairs. Each threshold is fixed across

K

. The rule and ReAL columns retain their measured NFE. The corresponding cap-sweep plot appears in Appendix C .

$K$	$q_{\rm seg}$	Rule	$\Delta$ NFE	$\Delta$ IR	Paired SE	95% CI
12	2.6	3:12	-0.35	-0.009	0.046	[-0.102, +0.082]
12	3.8	2:12	-0.68	-0.155	0.073	[-0.305, -0.022]
16	2.6	3:16	+0.76	+0.152	0.058	[+0.040, +0.268]
16	3.8	2:16	+0.16	+0.056	0.075	[-0.097, +0.195]
24	2.6	3:24	+1.83	+0.621	0.073	[+0.481, +0.765]
24	3.8	2:24	+1.41	+0.704	0.088	[+0.531, +0.876]

Appendix

Table 11: FLUX cap comparisons. Differences are ReAL minus rule. Positive

\Delta

NFE means more compute. Intervals are exploratory and unadjusted for operating-point selection.

$T$	Method	Setting	NFE	IR	IR ratio (%)
15	Dense	N/A	15.000	0.743	100.0
15	Rule	3:6	6.000	0.047	6.3
15	Rule	4:6	7.000	0.302	40.6
15	ReAL (guided)	$q_{\rm seg}=3.2$	6.245	0.198	26.6
15	ReAL (guided)	$q_{\rm seg}=2.2$	6.730	0.312	41.9
15	ReAL (guided)	$q_{\rm seg}=1.65$	7.265	0.382	51.4

Appendix

Table 12: Completed SD3-Medium evaluation with 200 DrawBench prompts. Each step count has its own dense reference.

Method	Setting	NFE	IR	CLIP	LPIPS $\downarrow$	PickScore
Dense	N/A	28.00	0.736	27.48	0.000	22.06
Rule	6:3	14.00	0.678	27.39	0.096	21.98
Rule	4:4	11.00	0.598	27.40	0.182	21.85
ReAL (guided)	$q_{\rm seg}=0.9$	14.76	0.678	27.47	0.105	21.93
ReAL (guided)	$q_{\rm seg}=1.25$	12.43	0.594	27.49	0.171	21.88
ReAL (guided)	$q_{\rm seg}=1.65$	10.84	0.580	27.61	0.214	21.86

Appendix

Table 13: SD3-Medium, DrawBench-100. TeaCache uses output-delta reuse. TaylorSeer is an output-level reimplementation.

Method	Setting	NFE	IR	CLIP	LPIPS $\downarrow$	PickScore
Dense	N/A	28.00	0.879	27.78	0.000	22.96
Rule	3:6	8.00	0.755	27.95	0.182	22.57
Rule	4:6	9.00	0.800	27.93	0.152	22.60
Rule	4:4	11.00	0.830	27.86	0.103	22.77
ReAL (guided)	$q_{\rm seg}=0.6$	11.87	0.820	27.89	0.098	22.81
ReAL (guided)	$q_{\rm seg}=1.25$	8.25	0.794	27.93	0.237	22.67

Appendix

Table 14: SD3.5-Large, DrawBench-100, with the same implementation labels as Table 13 .

Method	Setting	NFE	IR	CLIP	LPIPS $\downarrow$	PickScore
Dense	N/A	30.00	0.546	26.25	0.000	22.40
Rule	6:3	15.00	0.486	26.34	0.034	22.26
Rule	2:6	8.00	0.140	26.17	0.189	21.76
Rule	3:6	9.00	0.221	26.42	0.150	21.89
ReAL (guided)	$q_{\rm seg}=0.35$	14.29	0.405	26.41	0.070	22.16
ReAL (guided)	$q_{\rm seg}=0.9$	9.27	0.222	26.18	0.170	21.94

Appendix

Table 15: Lumina-Next-SFT, DrawBench-100. Quality and NFE compare adaptive advancement with fixed-stride and dense-head allocation.

Method	Setting	NFE	IR	CLIP	LPIPS $\downarrow$	PickScore
Dense	N/A	28.00	1.065	N/A	N/A	N/A
Rule	6:3	14.00	1.032	N/A	N/A	N/A
Rule	4:4	11.00	0.976	N/A	N/A	N/A
ReAL (guided)	$q_{\rm seg}=1.25$	12.66	1.037	N/A	N/A	N/A
ReAL (guided)	$q_{\rm seg}=1.65$	11.14	1.013	N/A	N/A	N/A
ReAL (guided)	$q_{\rm seg}=3.2$	9.37	0.905	N/A	N/A	N/A

Appendix

Table 16: SD3-Medium on the 150-prompt PartiPrompts subset. The 28-step dense reference belongs to this prompt set.

Method	Setting	NFE	IR	CLIP	LPIPS $\downarrow$	PickScore
Dense	N/A	50.00	1.016	27.59	0.000	23.17
Rule	2:12	7.00	0.831	27.26	0.275	N/A
Rule	3:12	8.00	0.893	27.57	0.222	N/A
Rule	3:8	10.00	0.963	27.54	0.163	N/A
ReAL	$q_{\rm seg}=2.6$	7.65	0.884	27.46	0.300	N/A
ReAL	$q_{\rm seg}=3.8$	6.32	0.676	27.40	0.361	N/A

Appendix

Table 17: FLUX NFE sweep on DrawBench-100. Quality uses this sweep’s dense reference.

Policy	$q_{\rm seg}$	NFE	IR	Accepted $k$	NFE/iteration
Always carry	1.65	9.48	0.872	6.19	1.06
Fallback-only carry	2.6	10.14	0.814	10.13	1.96
Fallback-only carry	3.6	10.00	0.793	10.66	1.89
Carry if $k\leq 4$	1.65	11.06	0.863	8.73	1.73

Appendix

Table 18: Carry-policy study on FLUX, 100 prompts. Always-carry and short-only carry use

q_{\rm seg}=1.65

to isolate cross-segment reuse. Fallback-only rows report additional combinations of thresholds and policies.

Figure 8: Adaptive allocation across segment caps, schedules, and prompts. a , FLUX cap sweep. Marker shape identifies

K

, and color identifies the method. The rule uses

m\in\{2,3\}

and

k=K

. b , Expanded 200-prompt SD3-Medium sweep at

T=15

, with fixed-rule configurations shown at their actual NFE. c , On 150 PartiPrompts, the controller allocates more evaluations to the 98 prompts outside the Basic or Simple Detail categories than to the remaining 52. Each curve sweeps the segment threshold, and the prompt groups use the benchmark categories.

Figure 9: An empirical fixed-stride relationship across the tested backbones. (a) , Pooled power-law fit to 22 configuration means. (b) , Predictions after fitting on the other three backbones. Legend ratios summarize measured LPIPS divided by predicted LPIPS as geometric means within each held-out backbone. The largest individual relative prediction error is about 20.0% for Lumina. The fit summarizes configuration means that share prompts and dense references.

Figure 10: ReAL adapts evaluation placement. a , Mean LPIPS after dense continuation from a structured perturbation. Bars show one standard error across prompts. b , Mean terminal drift divided by mean summed dense-path discrepancy on SD3-Medium (

n=100

). Axis labels show controller NFE. The ReAL replay adds one unused terminal query on 42 trajectories. c , Mean selected advance over recorded regular starts in six trajectory bins (

n=100

). Tail steps are excluded.

Figure 11: Additional accelerator comparisons. The same prompts and noise compare ReAL with the original baseline set, including SpeCa, TaylorSeer, DPCache, Spectrum, JiT, and FlowCast. Runtime environments are specified above.

Figure 12: Complete HunyuanVideo qualitative comparison. Rows show dense sampling, direct step reduction, twelve acceleration baselines, and ReAL. Each prompt group displays frames 0, 32, and 64 from the same generated video. The panels retain the full frame and original colors.

Figure 15: Lumina-Next-SFT qualitative comparison. Dense sampling, ReAL, fixed stride, and a dense-head schedule. The fixed-stride and dense-head controls compare allocation strategies on the common prompts.

Flow Matching

Flow matching learns continuous-time vector fields that transport simple noise distributions into data for efficient generative modeling.

See all 663 papers

Oct 17, 2026

CorrFlow: Self-Correcting Flow Matching

Flow-matching models generate samples through iterative velocity predictions, but typically do not revisit these predictions using the intermediate representations that produced them. We introduce CorrFlow, a self-correcting flow-matching framework that reuses this information through a lightweight refiner. Under rectified-flow interpolation, the error of the one-step clean prediction equals the realized velocity residual scaled by the noise level. Motivated by this identity, the refiner learns to predict residual corrections from the base model's clean prediction, conditioning, noise level and hidden states from the same forward pass. The correction is added to the base velocity before each sampling update, so subsequent predictions are made from the corrected state. This requires no additional base-model evaluations and only one lightweight refiner evaluation per step. Our theoretical analyses establish conditions under which feature reuse outperforms raw-input affine predictors and refiner training achieves greater residual-risk reduction than continued base-model training under a matched training compute budget. Experiments on class-conditional image generation, text-to-video generation and image-to-video generation show consistent improvements over the corresponding base models, including gains in visual quality and physical consistency. Project Page: https://fogradio.github.io/corrflow-project

Hangwei Zhang, Nathan Rouillé, Chuanxia Zheng

PVG, Nanyang Technological University · Télécom Paris, Institut Polytechnique de Paris

Flow Matching Flow-Based Generative Modeling Conditional Image Generation

Ask PDF

May 12, 2026

Sharpen Your Flow: Sharpness-Aware Sampling for Flow Matching

Flow matching models generate samples by numerically integrating a learned velocity field, with each integration step requiring a neural network evaluation. Fast generation therefore requires using a small fixed evaluation budget effectively: the key question is not only how to integrate the flow, but where the sampler should spend its steps. We propose SharpEuler, a training-free sampler that profiles a pretrained model offline by estimating where the learned velocity field changes most rapidly along calibration trajectories. This finite-difference estimate defines a solver-aware sharpness profile, which is smoothed and converted by a quantile transform into a timestep grid for any desired inference budget. At test time, sampling remains ordinary Euler integration with the same number of model evaluations as a uniform schedule. We justify SharpEuler using three principles: a numerical principle identifying trajectory acceleration as the leading source of Euler discretization error, a variational principle deriving sharpness-based power-law timestep densities, and a statistical guarantee showing that the finite-sample calibrated sampler is stable at the terminal distribution level. Our experiments show that SharpEuler improves sample quality at fixed budgets, reducing inter-mode leakage and increasing mode coverage.

Aditi Gupta, Soon Hoe Lim, Annan Yu +1

Lawrence Berkeley National Laboratory · International Computer Science Institute · 1Lawrence Berkeley National Laboratory, 2International Computer Science Institute +4

Flow Matching Flow-Based Generative Modeling

Ask PDF

Feb 5, 2026

Stable Velocity: A Variance Perspective on Flow Matching

While flow matching is elegant, its reliance on single-sample conditional velocities leads to high-variance training targets that destabilize optimization and slow convergence. By explicitly characterizing this variance, we identify 1) a high-variance regime near the prior, where optimization is challenging, and 2) a low-variance regime near the data distribution, where conditional and marginal velocities nearly coincide. Leveraging this insight, we propose Stable Velocity, a unified framework that improves both training and sampling. For training, we introduce Stable Velocity Matching (StableVM), an unbiased variance-reduction objective, along with Variance-Aware Representation Alignment (VA-REPA), which adaptively strengthen auxiliary supervision in the low-variance regime. For inference, we show that dynamics in the low-variance regime admit closed-form simplifications, enabling Stable Velocity Sampling (StableVS), a finetuning-free acceleration. Extensive experiments on ImageNet

256\times256

and large pretrained text-to-image and text-to-video models, including SD3.5, Flux, Qwen-Image, and Wan2.2, demonstrate consistent improvements in training efficiency and more than

2\times

faster sampling within the low-variance regime without degrading sample quality. Our code is available at https://github.com/linYDTHU/StableVelocity.

Donglin Yang, Yongxing Zhang, Xin Yu +5

University of Hong Kong, Hong Kong · University of British Columbia, Canada · Vector Institute for AI, Toronto, Canada +2

Flow Matching Image Generation Flow-Based Generative Modeling Variance Reduction Diffusion Model Inference Acceleration

Ask PDF

Abstract

Flow Matching

Explore similar work

CorrFlow: Self-Correcting Flow Matching

Sharpen Your Flow: Sharpness-Aware Sampling for Flow Matching

Stable Velocity: A Variance Perspective on Flow Matching