Multiview fine-tuned DiT checkpoints
Thank you guys so much for the release, it’s def the best image to 3d model released so far, and making it open source is really amazing. I would really like to contribute and would kindly request the release of the finetuned DiT safetensors (dense + sparse512 + sparse1024 on the paper branch, or the Pixal3D-T equivalents). I've checked all released weights, and the multiview fine tuned DiTs do not appear to be available anywhere. Since the conditioner needs no weights at all, multiview support should come from finetuning the singleview model on 2–6 randomly sampled views. So the only missing component for multiview inference would be the finetuned DiT. Would you consider releasing those fineuned DiT weights? Since the Multiview Dino encoder is already in the repo and the conditioner is weight free, the community could likely contribute the inference wiring via PR. The weights are the only blocker for reproducing Table 3, and given the paper they are already available. If this is not possible for any reason, please let me know in this case I would fine tune it myself.