The quality of image generation models trained with diffusion or flow matching is largely fixed by the training budget: at inference, the only available lever is the number of sampling steps, and additional steps improve quality only up to a point. We present a training recipe, Scout-Refine-Correct (SRC), that instead gives a flow-matching generator a genuine inference-time refinement capability. At each sampling step the model takes a scout rollout -- K steps ahead along its own trajectory -- extrapolates a full-sample estimate, and feeds that estimate back to itself as a self-conditioning signal before taking the real step. This can be repeated for R refinement rounds. Two correction losses, applied during training at states the model's own rollouts actually reach rather than only on the ground-truth path, teach the model to use this signal to reach new samples instead of collapsing to a suboptimal one. On ImageNet-256, SRC improves LightningDiT-B from FID14.5 to FID7.26 at NFE550, matching a privileged best-of-N baseline that uses an oracle FID verifier to choose among samples. SRC uses no external verifier, no reward signal, and, after ablation, almost no extra parameters over the base architecture. Combining SRC with best-of-N improves the result further, to FID3.81 at NFE~4400, showing that the two axes are complementary. We ablate the main design choices -- the architectural pathway for the scout signal, the scout horizon, each correction loss, and the training schedule to demonstrate key working components of training.
Figures & tables
Figure 2: In-place refinement of a basic self-conditioning model (no scout, no correction loss): FID vs. refinement rounds R . Quality saturates at ≈ 13.3 almost immediately – extra refinement compute buys nothing without the correction losses of Sec. 3.2 .
symbol
meaning
value
ρ
self-cond fraction of batch
0.5
δ
scout step size (fixed)
0.02
Kmin..Kmax
scout-count range
1..3
N
re-noise draws per rollout point
2
λsc,λm
loss weights
1,1
Table 1: Training recipe hyperparameters.
Figure 3: FID vs. NFE on ImageNet-256: sampling-step scaling alone (base model, flat), vanilla self-conditioning R -scaling (plateaus at ≈ 13.3), and SRC’s single-sample K×R frontier (global best FID 7.26 at NFE 550).
NFE
SRC ( K=2,R=1 )+Bo N
baseline Bo N
200
7.89
8.25
400
6.53
6.96
800
5.45 ( N=4 )
5.97 ( N=8 )
1600
4.12
-
Table 2: Same-budget comparison: SRC anchored at K=2,R=1 versus the baseline best-of- N oracle
N
SRC ( K=4,R=2 )+best-of- N
NFE
1
7.26
550
2
5.99
1100
4
5.02
2200
8
3.81 (new global best)
4400
Table 3: Scaling SRC+best-of- N past where the baseline best-of- N oracle was ever measured.
Figure 4: SRC K/R refinement compared with scaling the number of inference steps and with scaling the classifier-free guidance weight.
Figure 5: (a) FID vs. inference-time K at R=1 . (b) FID at four decisive operating cells. main-corr is the single most critical term; scout-corr is also clearly load-bearing, while K -conditioning doesn’t introduce significant gains.
R=2
R=3
with self-conditioned FM
0.013
0.003
without self-conditioned FM
0.164
0.183
Table 4: Mean LPIPS between refinement round R and round R+1 , averaged over images and sampling steps. The two rows are the self-conditioning baseline that keeps the term and the later model that drops it.
Figure 6: Pixel differences for SRC refinement and for classifier-free guidance scaling, on the same examples.
Kmax
best FID
2
7.21
3
7.18
4
7.49
6
8.30
8
9.99
Table 5: Training-time scout horizon ( Kmax ). Kmax=3 is the best of the tested values.
(a,b,c) steps
FID @ best KR
(120k, 0, 90k) – no stage b
7.85
(120k, 90k, 0) – no stage c
8.88
(120k, 60k, 30k)
7.18
(120k, 30k, 60k)
7.53
(90k, 90k, 30k)
7.34
Table 6: Training-stage budget. The (120k,60k,30k) split is the best of the five schedules tested.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: The architecture used from the self-conditioning baseline through SRC: s enters only as concatenated tokens from a bias-free embedder, and one AdaLN slot is driven by time and class. K is not a network input.
Figure 8: Matched-seed comparison of SRC refinement and vanilla self-conditioning refinement. Left: SRC at NFE {50,100,200,550} . Right: vanilla self-conditioning at the same NFEs (reference, R=1 , R=3 , R=10 ). The right-hand family is the baseline that plateaus in Fig. 2 .
Flow-based generative modeling learns probability transports from simple noise to data for fast sampling, density estimation, and controllable generation.
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
We aim to improve frozen flow-matching image generators by adding inference computation inside the denoiser, without changing model weights or the outer sampler. Existing generators usually spend extra test-time computation by increasing the number of sampling steps, which repeatedly evaluates the entire denoiser and couples quality gains to sampler cost. A key challenge is how to use extra computation inside a frozen transformer denoiser: the method must decide which tokens, layers, and sampling times receive repeated updates while preserving the original generation pipeline. We introduce a training-free looping framework that repeatedly applies selected transformer layers inside each denoising call. Dense and Sparse Token Loop vary the token scope; Sampling-Progress Gating and the loop layer range specify when and where looping is active; loop count and strength control the repeated updates; and Loop Guidance combines ordinary and looped vector-field predictions. Across two Scale-RAE model scales, loop variants improve primary and auxiliary quality metrics with competitive quality--efficiency trade-offs. Loop Guidance further improves both primary metrics across all three tested models; on Scale-RAE DiT2.4B, it raises GenEval from 0.4471 to 0.5691 and DPG-Bench from 0.7656 to 0.8053. Code will be released.
Yuanyi Yan, Xinzhe Rao, Canyu Shen +5
Huazhong University of Science and Technology · Tongji University · King Abdullah University of Science and Technology +2
We propose continuous adversarial flow models, a type of continuous-time flow model trained with an adversarial objective. Unlike flow matching, which uses a fixed mean-squared-error criterion, our approach introduces a learned discriminator to guide training. This change in objective induces a different generalized distribution, which empirically produces samples that are better aligned with the target data distribution. Our method is primarily proposed for post-training existing flow-matching models, although it can also train models from scratch. On the ImageNet 256px generation task, our post-training substantially improves the guidance-free FID of latent-space SiT from 8.26 to 3.63 and of pixel-space JiT from 7.17 to 3.57. It also improves guided generation, reducing FID from 2.06 to 1.53 for SiT and from 1.86 to 1.80 for JiT. We further evaluate our approach on text-to-image generation, where it achieves improved results on both the GenEval and DPG benchmarks.