Latent Diffusion Models dominate generative modeling but face two-stage training complexity and the inherent reconstruction limits of pre-trained autoencoders. Pixel-space Flow Matching offers a simplified end-to-end alternative yet often encounters training difficulties when using the standard v-prediction objective. Recent research attributes this difficulty to the off-manifold nature of noised targets, motivating a shift toward x-prediction. In this paper, we investigate this explanation through patch-wise Principal Component Analysis (PCA) and identify a bandwidth bottleneck. The velocity in minor components is largely determined by the raw noisy input, requiring deep backbones to preserve information that competes with principal-component modeling. Through controlled experiments and representation analysis, we further examine why x-prediction remains effective: its parameterization provides a raw-input bypass, while its learned input representation can spontaneously concentrate on principal directions. Under low-rank observations, our analysis of the induced flow characterizes x-prediction as an implicit conditional-mean decoder, reconstructing discarded components from the retained information. These observations motivate the Raw-injected Transformer (RiT), which combines a lightweight raw-conditioned decoder with an output-side shortcut to separate semantic modeling from raw-state propagation. Our method restores effective v-prediction and achieves competitive performance on ImageNet 256×256, providing a perspective on the interaction between prediction objectives and architectural bandwidth.
Figures & tables
Figure 1: Qualitative results and quantitative comparison. The visual examples illustrate the generation quality of our model, while the table compares FID-50K across different model scales on ImageNet 256×256 . RiT-B and RiT-L are trained for 600 epochs, whereas RiT-H is trained for 480 epochs. All JiT [ 30 ] baselines are trained for 600 epochs. Our models consistently achieve lower FID-50K scores, even with fewer training epochs at the H scale.
Figure 2: Overview of RiT and the bandwidth bottleneck. Standard v -prediction must retain noisy input while learning semantic features, creating competing demands on backbone bandwidth. Enlarging the backbone can alleviate this conflict without extra raw injection, but incurs substantial computational overhead. RiT instead uses a lightweight raw-conditioned nonlinear decoder with sufficient post-injection capacity and an output-side shortcut. Details of the final layers are given in Section 4.5 . In contrast, x -prediction provides an inherent raw-input bypass and can learn compact representations concentrated on principal directions. Under low-rank conditions, its induced flow reconstructs discarded components as conditional means without additional decoder layers.
v -prediction
x -prediction
PCA Dim k
128
192
256
384
768 (Full)
128
192
256
384
768 (Full)
FID-50K ( ↓ )
33.99
36.49
38.75
47.41
149.07 (Fail)
35.01
37.56
39.58
40.83
40.39
Table 1: Effect of PCA projection. While v -prediction fails in the full 768-dimensional pixel space, restricting both data and noise to lower-dimensional principal subspaces not only restores effective training but even allows v -prediction to outperform x -prediction at k=128,192,256 . Its performance then deteriorates as additional minor directions are introduced.
Figure 3: Linear reconstruction fidelity ( R2 score) of xt from hl . We observe that v -prediction tends to retain substantially more noisy-input information in the residual path than x -prediction, consistent with the raw-state contribution required for minor-component velocity estimation. When the patch and hidden dimensions are comparable, preserving this 768-dimensional state can compete with the capacity needed for principal-component modeling. In contrast, the PCA-128 input has fewer stochastic dimensions, allowing high input retention alongside effective generative learning.
Figure 4: Eigenvalue distribution of image patches. Eigenvalues of the centered 768×768 patch covariance on the ImageNet 256×256 training set, shown in descending order. Their magnitudes span more than six orders, reflecting a strongly anisotropic distribution. Minor directions contain little data variance but receive unit-variance Gaussian noise in full-space training. Predicting their velocity therefore requires access to the noisy state even when the remaining regression error is small, motivating a raw-input shortcut that reduces this transmission burden.
Figure 5: Evolution of the learned input embedding. Left: effective rank of W=W2W1 , computed from the entropy of its normalized singular values to characterize the dimensionality of the learned representation. Right: energy in the first 280 PCA directions. Full-space x -prediction concentrates on principal directions, whereas v -prediction retains a broader spectrum. PCA-128 denotes projected data and noise, not an embedding bottleneck.
Intervention
Backbone width
Patch
Decoder width
Raw input
FID-50K ( ↓ )
Vanilla B
768
16
–
No
149.07
Vanilla L
1024
16
–
No
33.43
Vanilla B, smaller patches
768
8
–
No
28.70
Two-block decoder
768
16
128
Yes
220.78
Two-block decoder
768
16
256
Yes
34.68
Two-block decoder
768
16
384
No
186.16
Table 2: Interventions on bandwidth and raw access. All rows use v -prediction, 250K updates, and no guidance. Vanilla models have no low-rank input bottleneck. Decoder models use a rank-128 backbone input bottleneck and patch-8 decoder tokens, without an output-side shortcut. “Raw input” denotes an additional decoder input, not the ordinary backbone input.
Output shortcut
Objective
FID-50K ( ↓ )
None (standard decoder readout)
v
31.85
Basic learned gains
v
28.27
Free fractional, shared denominator
v
28.33
Anchored fractional, separate denominators
v
28.51
Anchored fractional, shared denominator (RiT)
v
27.97
Anchored fractional, shared denominator
x
28.49
Table 3: Output-shortcut ablation. All models use the B-scale backbone, a rank-128 input bottleneck, and a raw-conditioned two-block width-384 decoder. Results are measured after 250K updates without guidance. The output-side shortcut improves FID over the standard decoder readout, with the anchored fractional parameterization performing best among the tested variants. The x -prediction row uses the same architecture with the clean-data objective to compare the two prediction targets.
Bottleneck Width
128
256
384
768
FID-50K ( ↓ )
27.97
29.63
29.70
30.45
Table 4: Hyper-parameter ablation. Following the x -prediction baseline [ 30 ] , we find that a lower-rank input bottleneck and a high-noise sampling bias consistently enhance performance. All results are obtained with RiT-B after 200 epochs without classifier-free guidance.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Principal-component reconstruction. Original image (left) and its reconstruction using the leading 128 of 768 principal components (right). Much of the visible structure is preserved, illustrating how retained principal components can provide sufficient information for conditional completion, although some fine details may remain ambiguous.
Setting
RiT-B/16
RiT-L/16
RiT-H/16
Image resolution
256×256
256×256
256×256
Backbone patch size
16
16
16
Backbone depth
12
24
32
Backbone hidden width
768
1024
1280
Backbone attention heads
12
16
16
Input bottleneck width
128
128
256
Appendix
Table 6: Final model configurations. All final models use full pixel-space data and noise, a raw-conditioned decoder, and the shared-denominator anchored shortcut. Epoch counts are approximate. PCA projection is used only in the diagnostic experiments, rather than in the final models.
Variant
gr
gh
(γ,β)
Basic
1+c1
1+c2
(c3,c4)
Free fractional
(1+c1)/f2
(1+c3)/f2
(c4,c5)
Anchored fractional
n1/d2
n3/d2
(c4,c5)
Anchored, separate denominators
n1/d2
n3/d4
(c5,c6)
Appendix
Table 7: Coefficient assignments for the output-shortcut ablation. All operations are channel-wise. The final RiT uses the anchored fractional variant with a shared denominator, while retaining separate learned numerators for the raw and semantic branches.
Model
Updates
Epochs
EMA
CFG scale
CFG interval
RiT-B/16
250K
200
0.9996
3.1
[0.0,0.89]
RiT-L/16
250K
200
0.9996
2.5
[0.0,0.89]
RiT-H/16
250K
200
0.9996
2.4
[0.0,0.89]
RiT-B/16
750K
600
0.9996
2.8
[0.0,0.89]
RiT-L/16
750K
600
0.9996
2.6
[0.0,0.89]
RiT-H/16
600K
480
0.9996
2.35
[0.0,0.90]
Appendix
Table 8: Inference configurations for the reported guided results. Epoch counts are approximated from training updates and batch size. All configurations use 50-step Heun sampling with a final Euler step. Unguided ablations use w=1 and EMA 0.9996 with the same integration procedure.