Training-Free Hidden-State Refinement for Flow-Matching Image Generators

Yuanyi Yan

Manuscript in preparation

Paper (coming soon)Code

Teaser: qualitative gains and quality-efficiency trade-offs of token looping — Left: samples without looping, with Dense Token Loop, and with Loop Guidance. Right: looping improves GenEval and DPG-Bench at a better quality-efficiency trade-off than simply adding sampling steps.

Qualitative Comparisons

Matched-prompt samples on Scale-RAE Qwen7B / DiT9.8B (seed 42): w/o Loop vs. Dense Token Loop vs. Loop Guidance.

GenEval prompts

DPG-Bench prompts

Method

Dense / Sparse Token Loop — replay a window of transformer layers for all tokens or only a selected subset, concentrating extra compute where it matters most.
Sampling-Progress Gating — activate looping only inside a mid-trajectory interval where refinement is most effective.
Loop count and strength — control how many repeated updates are applied and how strongly each update moves the hidden state.
Loop Guidance — combine ordinary and looped vector-field predictions, analogous to classifier-free guidance, for further gains on both primary metrics.

Results

Main results across three RAE-based generators

Model	Method	GenEval ↑	DPG-Bench ↑	ImageReward ↑	Time (s) ↓
Scale-RAE DiT2.4B	w/o Loop	0.4471	0.7656	0.0994	1.31
	Dense Token Loop (Ours)	0.5422	0.8007	0.4816	2.18
	Sparse Token Loop (Ours)	0.5082	0.7893	0.2798	1.69
Scale-RAE DiT9.8B	w/o Loop	0.5321	0.8003	0.4609	1.69
	Dense Token Loop (Ours)	0.5977	0.8178	0.7211	4.04
	Sparse Token Loop (Ours)	0.5857	0.8206	0.6365	2.44
RAEv2 SigLIP2-B (615M)	w/o Loop	0.3829	0.7131	-0.3385	0.80
	Dense Token Loop (Ours)	0.3795	0.7226	-0.2538	1.10
	Sparse Token Loop (Ours)	0.3815	0.7152	-0.2931	1.36

Loop Guidance improves both primary metrics on all models

Model	GenEval (w/o Loop)	GenEval (+ Loop Guidance)	DPG-Bench (w/o Loop)	DPG-Bench (+ Loop Guidance)
Scale-RAE DiT2.4B	0.4471	0.5691	0.7656	0.8053
Scale-RAE DiT9.8B	0.5321	0.6432	0.8003	0.8264
RAEv2 SigLIP2-B (615M)	0.3829	0.4532	0.7131	0.7521

Looping also beats the obvious alternative of adding sampling steps: a 25-step Dense Token Loop reaches 0.5422 GenEval and 0.8007 DPG-Bench in 2.18s, outperforming the 50-step no-loop baseline on quality at lower latency.

Citation

@misc{yan2026loopedflow,
  title  = {Training-Free Hidden-State Refinement for Flow-Matching Image Generators},
  author = {Yan, Yuanyi},
  year   = {2026},
  note   = {Manuscript in preparation}
}