Request: MatDiff & MatVAE pretrained checkpoints for academic use

[HaoRayx]

First of all — congratulations on the CVPR 2026 Highlight! I've been reading
your paper (arXiv 2512.17302) and the released code, and I really like the
MatVAE + flow-matching MatDiff design.

I'm a researcher at [Wuhan University] working on physically-based
rendering material generation. Specifically, I'm building a benchmark / study
of physical validity of generated PBR materials (energy conservation,
Fresnel consistency, channel-wise constraints), and I'd like to reproduce and
report MatLat as a strong, state-of-the-art baseline in my experiments.

I've already cloned the repo and prepared the parts I can obtain myself:

• stabilityai/stable-diffusion-3.5-medium (HF, gated)
• RealESRGAN_x2plus.pth and big-lama.pt (links in README)
• null_prompt_embeds.pt / null_pooled_prompt_embeds.pt (generated via the
  provided cache commands)

The two checkpoints I could not find a download link for are the trained
MatLat weights themselves:

• checkpoints/matvae_control_ema.pt
• the trained MatDiff weights (--weights path/to/matdiff_weights)

Could you share these (or point me to where they'll be released)? I'm happy
to use them strictly for non-commercial academic research, will not
redistribute them, will cite the paper, and will sign a data-use
agreement if needed.

Thanks again for open-sourcing the code — really looking forward to the
weights!

Best,
[HaoRay] — [HaoRay@whu.edu.cn]

[32V]

Hi HaoRay,

Sorry for the delayed response, and thank you for your interest in MatLat. We have just released the pretrained MatVAE and MatDiff checkpoints (paper version). Please check the latest repository update and follow the instructions in the README to download and use them.