Work

Luminance-Preserving Cascade: DecouplingStructure and Color in Two-Stage Flow Matchingfor Image and Video Generation

Year: 2026
Type: preprint
Abstract: We present a two-stage flow-matching cascade that generates an image by first predicting its luminance and only then colouring it, and we measure exactly what that decomposition buys and what it costs.The architecture is a "luminance-preserving cascade". Stage 1 is a single-channel flow-matching model that transports Gaussian noise to the Rec. 709 luminance plane; Stage 2 is a two-channel model that predicts the chroma planes conditioned on Stage 1's output. Because the chroma is applied as an additive offset in YCbCr and decoded back to RGB, the luminance of the decoded image is identical to the luminance Stage 1 produced -- an algebraic identity, not a penalty term. Out-of-gamut chroma is projected onto the feasible set by a closed-form radial rescaling that also leaves the luminance untouched. The invariant was verified at 1.19e-07 (the float32 machine epsilon) for every sample, and it is bit-identical under a U-Net and under a diffusion transformer, because it is a property of the colour-space algebra rather than of any network.What the decomposition buys is a contract rather than a better image. The luminance is fixed before any colour decision is taken, so the colour controls cannot move it (measured: 2.4e-07 over a recolour sweep). Re-running Stage 2 under a different tone prompt re-colours an image in 24.5 ms per image while the luminance stays exactly where it was, and the same request against a joint RGB model would have to regenerate the frame. Illumination and tone are controlled independently by text, with a Spearman rank correlation of 0.998-0.999 for the illumination sweep, and the requested colour diversity is an explicit, strictly monotone channel (Spearman 1.000). For video, the same architecture interprets Stage 1 as a per-pixel motion-direction predictor: the photometric change between adjacent frames is measured to be a single global gain, and a grayscale-first Stage 1 predicts luminance better than a joint RGB one (MSE 0.040 against 0.050). What it costs is reported just as explicitly. Against a monolithic RGB model of matched capacity, budget, prompts and sampler, the cascade is behind on image quality (FID 14.3 against 9.5 at 32x32, 90.2 against 74.7 at 64x64), and the colour statistics localise the cause: the cascade keeps 77% of the data's colour diversity against the control's 87%, because sampling luma and chroma independently cannot correlate a sample's colour with the particular luminance it was handed. Several of our own hypotheses were refuted by our own measurements (luminance is not more temporally consistent than colour; the temporal hierarchy is not cheaper on a luminance state; Stage 2 needs four sampling steps, not one) and are reported as null results. Scope: CIFAR-10 at 32x32, Tiny-ImageNet at 64x64, and UCF-101 for video; the video baseline is an in-house monolithic model of identical architecture, data and budget, not a reproduction of a state-of-the-art system, and we report no FVD because the I3D features of the original metric cannot be obtained offline. All code, checkpoints and evaluation scripts accompany the paper. This is a preprint of a manuscript under submission to Springer Nature's Signal, Image and Video Processing. It has not been peer reviewed; the numbers may change in the final version. Please cite it as a preprint, and do not treat the absence of a journal DOI as a statement about quality.
Source: Figshare
Author: Guanyu Yue
Language: English
Open Access status: green
OpenAlex ID: W7236710257
Cites:
Cited by:
Related to: 9