Organizations: Institute of Information Engineering, Chinese Academy of Sciences · School of Cyber Security, University of Chinese Academy of Sciences · East China Normal University · Alibaba Group
Flat color-block drafts specify layout and regional color while leaving interior detail unspecified, yet pretrained flow-based generators can preserve their flat appearance even under photorealistic prompts. We trace this behavior to a cross-scale mechanism linking fine pixel variation to regional means in variational autoencoder (VAE) latent space. Flattening photo regions systematically shifts their latent means, with offsets concentrated along a shared channel direction within each backbone. Controlled field exchanges in FLUX.1-dev and Qwen-Image establish the effect of this displacement on texture generation: adding the photo-derived regional mean field to a flat latent restores texture, whereas adding its complementary field alone largely preserves flatness. A second-order encoding expansion explains how encoder curvature can convert zero-mean perturbations into mean shifts. An analytically solvable conditional Gaussian flow shows how regional noise averaging enables near-unit endpoint sensitivity at high signal-to-noise ratios. This mechanism motivates training-free rendering through pixel dither before encoding, mean-field scaling, and early boundary holding. Paired low-rank adaptation (LoRA) with draft conditioning improves content recovery and produces photo-like regional means under the standard velocity objective. Explicit regional mean guidance remains effective after adaptation, continuously controlling texture and improving color fidelity. These findings connect encoding shifts to persistent generation behavior and support interventions evaluated for texture, spatial fidelity, and semantic recovery.
Figures & tables
Figure 1: Flatness follows regional latent means. FLUX.1-dev with matched prompt/noise. Left: draft d and photo x , with equal regional RGB means before quantization. Their VAE latents zf,zp have different regional means, measured by A . In the output grid, rows select draft/photo mean fields and columns select draft/photo complementary fields ( b=0/1 ), using the soft-mask operator P . The photo complement alone leaves surfaces flat; the photo mean field restores texture. Matrix T : geometric texture ratios over 300 images. Curves: 100 images; dashed lines mark source texture T=1 and identity π=α .
Figure 2: Regional means connect encoding, trajectory retention, and rendering control. Pre-encoding dither produces a latent displacement; its mean field is scaled while its spatial complement is retained. The lanes represent the operational split P and I−P . Regional noise averaging motivates high mean retention. Boundary tokens are held early and then released; optional draft-conditioned LoRA and mean guidance act on the same velocity predictor. The guidance target uses δd=A(zd−zf) . Real FLUX.1-dev panels show one saved case: the draft (left), dither with mean scale 1.15 and boundary holding (upper right), and the flat-start rendering (lower right). These panels illustrate the training-free branch. Paths, fields, and coordinate insets are schematic; the gray path denotes a separate uncorrected run.
Figure 3: Training-free interventions with fixed prompt and noise within each comparison. (a) Living room: with FlowEdit, mean injection releases fine variation but retains outlined surfaces; dither gives a more photographic rendering. (b) Boat: with RF-Inversion, both mean injection and dither recover boat and water detail. (c) Skateboard: draft corrections retain the gray layout better than Tile. (d) Person: Tile recovers a recognizable subject, while draft corrections misinterpret the silhouette. “Mean” injects only the mean field; “+ hold” adds boundary holding to dither at scale 1.15. Tile uses strength 0.4. All panels retain the full frames.
Model
eˉM
eˉC
κ
FLUX
2.22
-0.20
0.96
Qwen
1.95
-0.20
0.90
Table 1: Photo-field effects and endpoint retention. Factorial: 300 FLUX and 100 Qwen images; dose regression: 100 per model. Effects are on log texture; κ is the terminal-position slope. Bold emphasizes the mean-field effect and retention.
Table 3: Draft interventions on 100 matched confirmation images per pipeline. Photo: photo-style rate. Means retain the flat complement; full dither changes both fields. Bold/underlining rank rendering scores within each pipeline; texture is unranked.
Family
Method
Overall rendering
Content recovery
Regional fidelity
S2↑
HPSv3 ↑
Classes ↑
Objects ↑
ΔE↓
Edges ↑
Frozen baselines
SDEdit 0.82
0.00
-5.38
0.36
0.34
2.5
0.40
SDEdit 0.92
0.07
2.72
0.38
0.29
6.0
0.14
ControlNet Tile 0.4
0.14
10.71
0.54
0.54
12.4
0.22
Training-free corrections
Dither
0.35
11.35
0.51
0.41
4.1
0.19
Dither + hold
0.33
9.94
0.51
0.44
3.3
0.33
Table 4: Photo rendering on the same 300 confirmation images. Methods are grouped by their use of paired draft/photo training. Tile 0.4 is selected on the separate tuning set. Dither + hold and LoRA mean guidance use scale 1.15. Values are point estimates. Bold: best; underline: second-best among the seven methods.
Figure 4: Mean-target control with LoRA, draft, prompt, and noise fixed. All five tested scales are shown. Increasing the target reveals surface detail alongside geometry changes. Blue marks the selected scale 1.15.
Figure 5: Regional correction and control on 300 images. (a) Regional RMS errors: text-only base and draft-conditioned LoRA. (b) Condition-dose response. Both probes use pure noise and guidance input one. (c) Pooled texture with fixed LoRA and guidance input 3.5; square: selected scale 1.15.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Subset
Images
Role
Direction fit
1,000
Shared direction and regression
Mechanism
300
Counterfactual trajectories
Confirmation
300
Renderer evaluation
Photo reference
1,000
Distributional distances
Development
48
Early development
LoRA pairs
2,221
Paired draft/photo training
Appendix
Table 5: Natural-image subsets. The first five subsets are disjoint. The LoRA subset excludes development, mechanism, confirmation, and reference images, and includes eligible direction-fitting images. The first 100 mechanism images support tuning; the original 48-image LoRA probe and ControlNet selection use subsets of that set. Final LoRA readouts use all 300 confirmation images.
Metric
What is measured
How to interpret it
Appearance diagnostics
Tgeom
Geometric mean of image-wise, source-relative interior high-pass variation.
Photo-field, boundary-control, Qwen, and extended quality statistic.
Tpool
Area-weighted regional variation pooled across images, then divided by the source pool.
Dither-field, cross-method, and LoRA control diagnostic.
Rendering: fidelity, content, and preference
ΔE↓
Area-weighted distance between regional mean Lab colors of output and draft.
Lower: closer regional colors.
Edges ↑
Draft-boundary recall within four pixels of an output edge, minus mismatched-boundary recall.
Higher: stronger alignment above chance.
Appendix
Table 6: How to read the experimental metrics. Arrows indicate the preferred direction for rendering scores. Texture measures detail relative to the source, so it carries no preference arrow. S1 and S2 are success rates defined for this study. Black rules separate comparison groups; numeric emphasis is defined in each table caption.
Kontext guidance 2.5; Redux guidance 3.5; original Canny control strength 0.7; 100-image subset.
Appendix
Table 7: Implementation settings. The guidance values in RF-Inversion, the generator, and regional mean feedback refer to different operations. Rows describe the executed protocols; the planned LoRA step count is distinguished from the checkpoint used for evaluation.
Measure
FLUX
Qwen
Offset energy along u
0.78
0.95
Synthetic/photo direction cosine
0.93
0.95
Half-response noise amplitude
0.0090
0.0119
Mean-field effect on log texture
2.22 [2.17,2.28]
1.95 [1.84,2.06]
Complement effect on log texture
-0.20 [-0.23,-0.16]
-0.20 [-0.27,-0.14]
Factorial interaction
0.20 [0.14,0.26]
0.49 [0.38,0.60]
Appendix
Table 8: Mechanism measurements. Encoder statistics use held-out regions from 300 photos and the synthetic block set. Half-response amplitude is interpolated in log amplitude at half the response to the largest tested amplitude (0.08). Factorial effects use 300 FLUX images and 100 Qwen images; controls and dose regressions use 100 per model. Brackets are 95% image-bootstrap intervals. A render is classified as flat when its texture ratio is below 0.3. Values are descriptive. Bold emphasizes encoder concentration, mean-field effects, and retention within each backbone.
Control
FLUX
Qwen
Photo-field log-texture gain
2.09 [1.99,2.20]
1.70 [1.59,1.82]
Random-field log-texture gain
-0.27 [-0.36,-0.17]
-0.06 [-0.16,0.05]
Object-only: object texture
0.74
0.95
Object-only: background texture
0.24
0.29
Shared direction only: texture
0.88
0.84
Shared direction only: color error
13.5
9.0
Appendix
Table 9: Direction, locality, and prompt controls on 100 mechanism images per model. All texture values are ratios to source-photo variation. Brackets are paired 95% image-bootstrap intervals for log-texture gains; dashes mark conditions not run.
Model
Start s
All
Small
Mid
Large
FLUX
0.96
0.67
0.71
0.62
0.67
FLUX
0.92
0.86
0.88
0.82
0.88
FLUX
0.87
0.96
0.97
0.95
0.97
FLUX
0.82
1.01
1.03
1.01
0.99
FLUX
0.71
1.01
1.04
0.99
0.99
Qwen
0.98
0.17
0.22
0.16
0.14
Appendix
Table 10: Endpoint retention on 100 synthetic layered layouts, stratified by region size. The median group sizes are 21, 145, and 1,236 latent cells. The logistic fit against the tested detectability predictor has R-squared 0.09 for FLUX and 0.17 for Qwen; an area-only fit gives 0.00 and 0.04. These are endpoint two-point responses, distinct from the multi-dose COCO regression. Bold marks the standard starting noise and its retention readouts.
Figure 6: Encoder response in FLUX.1-dev and Qwen-Image. Top: mean displacement along each model’s fitted photo direction as white-noise amplitude increases. Bottom: displacement for band-pass perturbations at amplitude 0.02. Points average the tested base colors; lines connect sampled settings and show no uncertainty bands. Fine-scale variation changes a regional latent average.
Table 17Table 18Table 19Table 20Table 21Table 22
Contrast
n
Outcome
Difference [95% CI]
Guided - alone
300
S2
0.10 [0.04, 0.15]
Guided - alone
300
ΔE
-3.15 [-3.40, -2.90]
Guided - alone
300
Edges
-0.02 [-0.02, -0.01]
Guided - alone
300
HPSv3
1.63 [1.39, 1.89]
Guided - alone
300
Objects
-0.02 [-0.05, 0.01]
Guided - full
300
S2
0.15 [0.09, 0.20]
Appendix
Table 25: Paired confirmation contrasts. Full denotes scaled dither with boundary holding; means denotes the dither-induced mean field. Each contrast resamples the same image IDs in its two conditions. Bold marks differences whose displayed intervals exclude zero, in either direction.
Flow-based generative modeling learns probability transports from simple noise to data for fast sampling, density estimation, and controllable generation.
We propose Perceptual Flow Matching (PFM), a simple yet effective framework for few-step generation in flow-matching models. Rather than performing velocity regression in the conventional VAE latent space, PFM supervises flow matching in a perceptual feature space using pretrained perceptual models. This simple change substantially improves the few-step generation capability of flow-matching models, reducing the number of sampling steps from 35-50 to 4-8 while preserving generation quality. Unlike existing acceleration and distillation approaches, PFM requires neither teacher models nor auxiliary score networks and can be integrated into standard flow-matching training pipelines with minimal modifications. Extensive experiments on image generation, video generation, and image editing tasks demonstrate that PFM consistently produces high-quality results while producing fewer artifacts than existing distillation-based methods. We further show that perceptual supervision shifts the regression minimizer from mean-seeking to mode-seeking, biasing predictions toward on-manifold modes that remain accurate under coarse few-step integration. Our results reveal that standard flow-matching training can naturally yield high-quality few-step generators when supervised in an appropriate representation space. We hope this insight inspires future research into representation-aware objectives for efficient generative modeling.
Chuyang Zhao, Yifei Song, Hongfa Wang +7
1Joy Future Academy · 2Fudan University · 3Tsinghua University +1
Most diffusion and flow-matching generators define the prior, probability path, and prediction target in the same representation space. Latent diffusion improves efficiency by moving this path into an autoencoder latent space, but the final sample is still produced by a separately trained decoder. This separation creates a mismatch: the generator is optimized for latent-space prediction, while final quality depends on how the decoder handles generated latents that may differ from clean encoder outputs. We introduce CrossFlow, a cross-space flow formulation that maps noisy latent inputs directly to pixel-space images. The key technical step is a velocity-free one-step objective: the latent trajectory defines the training path, but the supervised prediction is an image rather than a latent displacement. This lets one model act both as a one-step latent-to-pixel generator and as a decoder replacement for latent diffusion pipelines. On class-conditional ImageNet-1k at 256×256, CrossFlow-XL achieves 1.62 FID with one function evaluation. Ablations show that the latent encoder and pixel-space perceptual and adversarial losses are important for fidelity. These results indicate that cross-space flow objectives can combine the efficiency of latent representations with direct pixel-space supervision, without requiring a separate decoder at inference.
Xiyuan Wang, Xiao Zhang, Yang Li +4
Institute for Artificial Intelligence, Peking University · Tencent · Fudan University
Flow-based image generators achieve strong semantic synthesis but still produce localized structural distortions, such as malformed hands and limbs. Existing post-training methods typically optimize global scalar rewards, which provide weak spatial credit assignment and may alter well-formed regions while correcting local defects. We present Diagnose-to-Generate (D2G), a grounded alignment framework that separates visual diagnosis from generative optimization. We construct Achilles-600K, a region-annotated dataset for structural anomalies, and train ARGUS through supervised cold-start and Region-Aware Group Relative Policy Optimization to predict anomaly categories and bounding boxes. These diagnoses are converted into confidence-weighted spatial rewards for our Patch-level Flow-GRPO. We further identify that global stochastic exploration can shift semantic regions across group rollouts, making patch-wise relative advantages spatially inconsistent. Our Trajectory-Preserving Stochastic Sampling (TPSS) anchors early denoising with deterministic ODE trajectories and introduces stochasticity only at later steps, thereby limiting intra-group spatial drift while retaining local exploration. Experiments on FLUX.1-dev and SD3.5 show improved structural fidelity with preserved FID and CLIP scores. Consistent gains under HumanRefiner, zero-shot GPT-4.1, and blinded human evaluation further demonstrate that the improvements extend beyond the training reward model.
Yifan Li, Lishuai Gao, Tong Zhang +8
University of Science and Technology of China, Hefei, AnHui Province, China · Meituan Inc., Beijing, China