Flow-based generative models are trained on a hand-chosen prediction target, such as the velocity, the noise, or the clean data. This choice strongly affects training and sample quality. We introduce Anisotropic Flow Models (AnisoFlow), which replaces a fixed prediction target with a fully learnable one. AnisoFlow learns to scale the noise component of the target anisotropically, assigning a separate nonnegative weight to each direction in a jointly learned basis. A closed-form map recovers the velocity from the network's prediction, leaving the architecture, flow-matching loss, and sampler unchanged. Since recovery is exact, training optimizes the target purely to make it easier for the network to predict. We theoretically analyze the resulting prediction error and predict that the preferred noise weight in each direction scales near-linearly with data variance, in agreement with the weights learned empirically in pixel space. AnisoFlow improves FID from 1.90 to 1.55 over the JiT baseline on ImageNet 256×256, by learning only the prediction target, without representation alignment or architectural changes. In pixel-space text-to-image generation, AnisoFlow improves image quality and, through better-conditioned velocity recovery, largely removes patch-seam artifacts under reduced-precision inference.
Figures & tables
Figure 1: (a) High-resolution images generated by AnisoFlow, a pixel-space diffusion transformer with only 918M parameters trained from scratch, using a learned anisotropic prediction target. (b) By learning a well-conditioned prediction target, AnisoFlow removes the patch-seam artifacts produced by the x -prediction baseline under FP8 blockwise (W8A8) quantization ( DeepSeek-AI et al., 2025 ) .
Figure 2: AnisoFlow overview. (a) The network predicts the AnisoFlow target uW=Wϵ−x , where W is a learned positive semidefinite operator, and a closed-form map recovers the standard velocity. (b) In the eigenbasis of W , the target is parameterized separately in each direction, retaining the clean-data term −xi and applying a learned nonnegative weight wi to the noise term: uW,i=wiϵi−xi . Standard targets correspond to fixed weight profiles. AnisoFlow instead learns a nonnegative weight profile across directions.
Method
Noise weights
Noise basis
Init
FID ↓ @160ep
FID ↓ @600ep
Training sec/iter
Inference sec/step
JiT
Frozen (zeros)
–
–
2.68
1.90
0.803
0.179
AsymFlow
Frozen (binary)
Frozen (PCA)
PCA
2.34
1.76
0.803
0.179
AnisoFlow
Adam
Frozen (PCA)
PCA
2.25
1.63
0.805
0.180
AnisoFlow
Adam
Stiefel Adam
PCA
2.32
1.68
0.818
0.180
AnisoFlow
Adam
Spectral update
PCA
2.28
1.63
0.819
0.180
AnisoFlow
Adam
Stiefel Adam
–
2.28
1.60
0.818
0.180
Table 1: Effect of the learnable prediction target , isolating noise weights/basis optimizers and data-dependent initialization. JiT-H/16 network.
Figure 3: Emergent spectral structure in AnisoFlow. (a) Learned weights versus directional data variance biTΣbi , where Σ is the data covariance, for a frozen PCA basis and learnable bases with and without PCA initialization. The displayed range is trimmed to variances ≥0.015 . (b) Absolute inner products ∣vjTbi∣ between PCA directions vj and learned directions bi , in their native spectral order. Both initializations yield near-diagonal alignment, strongest at high-variance directions. All models are JiT-H/16; both learnable variants use the spectral update in Table 1 .
Method
Target
Params
GFLOPs
FID ↓
Hierarchical CNNs (U-Net-like)
ADM-G
ϵ
554M
2240
4.59
Hierarchical transformers (U-ViT-like)
RIN
ϵ
320M
668
3.42
SiD, UViT/2
ϵ
2B
1110
2.44
VDM++, UViT/2
ϵ
2B
1110
2.12
Table 2: ImageNet 256×256 pixel diffusion. Best reported FID, grouped by architecture family. FLOP estimation follows Yu et al. (2026) , and * marks the JiT evaluation protocol, which can give up to 0.08 lower FID than the ADM protocol. † : uses a pretrained representation encoder (e.g., REPA-style alignment or DINO-based losses) or self-supervised pre-training.
Figure 6
Figure 6: Qualitative text-to-image results. (a) 512×512 generations from AnisoFlow, the JiT-L/16-T2I baseline ( x -prediction), and the latent i1-1B model ( u -prediction), using identical prompts and sampling seeds. AnisoFlow renders finer texture than the x -prediction baseline (paper, fabric, waves, skin) and a sharper, more distinct style than the latent model, in line with its lower FID and pFID. (b) The same prompt and seed under 8-bit inference (INT8, FP8, and blockwise FP8). x -prediction develops patch seams or strong noise, while AnisoFlow remains stable.
Method
Params
OneIG ↑
Long Text * ↑
Gen Eval * ↑
DPG *↑
HPSv3 ↑
COCO-10k
FID ↓
pFID ↓
i1-1B (FLUX.2 VAE, u -pred.)
1.3B
0.511
0.853
0.826
86.3
7.11
62.3
62.4
JiT-L/16-T2I ( x -pred.)
917M
0.512
0.862
0.820
85.7
7.71
22.0
24.4
AnisoFlow-L/16-T2I (ours)
918M
0.516
0.864
0.814
85.8
8.03
20.1
22.7
Table 3: Text-to-image 512×512 . All models are trained using i1’s data, text encoder, and schedule, and tested using the same guidance and sampler settings. * indicates using i1’s rewritten prompts. Bold marks the best result within the matched pixel-space group.
Table 9
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Estimated prediction-error structure on ImageNet. (a) Time-averaged effective noise and data error scales versus PCA variance, with fitted scaling exponents. (b) Preferred weights estimated from the effective errors compared with the learned AnisoFlow weights. (c) Timestep dependence of the estimated preferred weights, normalized by the learned weights, compared with the schedule factor αt/σt (normalized by its geometric mean). Thin curves show individual directions, and the curve with markers shows the median. Results are shown for 18 PCA directions satisfying reconstruction-consistency criteria.
Noise basis
Init
@160ep
@600ep
Frozen (PCA)
PCA
3.1 / [0,0.76]
2.3 / [0,0.88]
Stiefel Adam
PCA
3.1 / [0,0.74]
2.1 / [0,0.90]
Spectral update
PCA
3.1 / [0,0.74]
2.3 / [0,0.88]
Stiefel Adam
–
4.4 / [0,0.62]
2.3 / [0,0.84]
Spectral update
–
3.6 / [0,0.66]
2.1 / [0,0.86]
Appendix
Table 6: Selected guidance settings for the AnisoFlow models in Table 1 . CFG scale / guidance interval selected by grid search for each model and training length.
Table 7: Text-to-image model and training settings. The 256×256 and 512×512 stages are used for the quantitative comparison, and the one-megapixel (1MP) and alignment fine-tuning (SFT) stages only for the qualitative samples in Fig. 1 . Each stage is initialized from the previous one.
Component
Additional state / operation
When applied
Directional weights
D scalars
Training and inference
Learned basis
D×D orthogonal matrix
Training and inference
Running spectral moment
D×D matrix
Training only
Basis transforms
Multiplication by B and B⊤
Each model evaluation
Directional recovery
Elementwise scaling
Each model evaluation
Spectral basis update
Eigendecomposition of a D×D matrix
Every training step
Appendix
Table 8: Additional computation and state introduced by AnisoFlow.
Figure 8: Effect of the JiT input bottleneck. Untrimmed version of Fig. 3(a) , showing all 768 directional variance/weight pairs for the frozen PCA basis and learnable bases with and without PCA initialization. Both learnable variants use spectral updates; the dashed line shows w=0.15λ . The near-linear profile at high variance changes noticeably in the low-variance directions beyond roughly the 256-dimensional patch-embedding bottleneck.
Figure 9: Cross-direction coupling. We leave 32 directional weights free to learn while freezing all complementary directions to either w=0 or w=0.15λ . Each point shows the final learned weight of one of these 32 free directions; gray crosses show the corresponding weights from a reference model in which all directions are learned jointly. (a) The free subset contains the 32 highest-variance PCA directions. (b) The free subset is stratified across the data spectrum. The high-variance free weights closely match the fully learned solution under both frozen backgrounds, while the lowest-variance directions show somewhat greater sensitivity to the complementary targets.
Flow-based generative modeling learns probability transports from simple noise to data for fast sampling, density estimation, and controllable generation.
Flow-based generation in high-dimensional pixel spaces is difficult because velocity prediction requires modeling high-dimensional noise, even when data has strong low-rank structure. We present Asymmetric Flow Modeling (AsymFlow), a rank-asymmetric velocity parameterization that restricts noise prediction to a low-rank subspace while keeping data prediction full-dimensional. From this asymmetric prediction, AsymFlow analytically recovers the full-dimensional velocity, leaving the network architecture and the standard flow-matching training and sampling procedures unchanged. On ImageNet 256×256, AsymFlow achieves a leading 1.57 FID, outperforming prior DiT/JiT-like pixel diffusion models by a large margin. AsymFlow also provides the first-ever route for finetuning pretrained latent flow models into pixel-space models: aligning the low-rank pixel subspace to the latent space gives a seamless initialization that preserves the latent model's high-level semantics and structure, so finetuning mainly improves low-level mismatches rather than relearning pixel generation. We show that the pixel AsymFlow model finetuned from FLUX.2 klein 9B establishes a new state of the art for pixel-space text-to-image generation, beating its latent base on HPSv3, DPG-Bench, and GenEval while qualitatively showing substantially improved visual realism.
Co-denoising pretrained representations such as DINO can substantially improve the training speed and quality of flow matching models, but it introduces a second denoising trajectory and requires carefully designed schedules. We propose a simpler alternative: predict the pretrained representation directly, then condition the model on its own prediction. This removes the need for a second ODE and any representation-specific denoising schedules, while retaining the benefits of representation guidance. Our approach converges substantially faster and achieves better generation quality as measured by FID score. On ImageNet, it outperforms the state of the art in latent space at 2x fewer epochs than prior methods; in pixel space, it improves FID over comparable prior methods by more than 20%. These results support a simple principle: do not denoise what you can predict. Our code is openly available at https://github.com/arijit-hub/dino_forcing.
Arijit Ghosh, Lucas Degeorge, Paul Couairon +3
ENPC, IP Paris · Ecole Polytechnique, IP Paris · AMIAD +1
In rectified-flow-based generative models, the neural network can be trained to predict two different targets, such as the instantaneous velocity or the data endpoint, to perform denoising. Although prior work shows that these parameterizations lead to different empirical behaviors, the mechanisms underlying their respective advantages remain to be underexplored, and how to combine them effectively is still unclear. In this work, we analyze how learning errors from different parameterizations affect the generation performance. We show that predicting the data endpoint has a clear training signal that stabilizes training, whereas predicting the velocity maintains stable sampling dynamics near the data manifold. Motivated by these insights, we propose Self-Consistent Flow (SC-Flow), a new method that unifies the benefits of both parameterizations. By employing a lightweight consistency loss, SC-Flow jointly trains a single network to predict both the local velocity and the data endpoint, and the consistency between the two predictions improves the model's performance. The method requires no major architectural changes and adds minimal computational overhead. Extensive experiments on image generation tasks demonstrate that SC-Flow substantially stabilizes optimization and improves the straightness of generation paths, leading to significant gains in generation quality over standard rectified-flow baselines.