Assess whether conversion improvements generalize to large-batch training

Determine whether the benefits of gain-matched ridge initialization and trunk learning-rate scaling for converting FLUX.2 Klein base 4B from latent space to pixel space survive at the global batch size of 128 used in prior conversion work.

Background

The paper finds that, at a global batch size of 8, combining gain-matched ridge initialization of the pixel input projection with a scaled trunk learning rate enables coherent image generation within 1,000 conversion steps. The authors explicitly note that they did not evaluate whether this advantage persists under the much larger global batch size of 128 used by Jiang et al. for related latent-to-pixel conversion experiments.

References

These probes used a global batch of 8 and 1K steps; we did not test whether the benefit survives at the global batch of 128 used by Jiang et al..

— Iris-3B: Going Beyond the Latent with Pixel-Space Diffusion Training, Conversion and Fine-Tuning  (2610.09450 - Cai et al., 7 Oct 2026) in Section 4.2, “Ridge initialization and trunk LR”