Request for ReconViaGen-v0.5 training details or minimal training recipe

Hello ReconViaGen authors,

Thank you for releasing ReconViaGen and the v0.5 inference code.
The multi-view fusion strategy built on top of TRELLIS.2 is very impressive and extremely useful for studying accurate multi-view 3D reconstruction.

I am currently trying to reproduce and further investigate the v0.5 pipeline for research purposes. I noticed that the v0.5 branch provides inference code, while the complete training pipeline/configuration is not included.

I completely understand that releasing the full training pipeline may require additional cleanup, especially considering the complexity of the TRELLIS.2-based training system.

Could you please consider sharing any of the following materials if possible?

1. Full v0.5 training code;
2. Training configuration files (DiT architecture, optimizer, schedule, dataset format);
3. The difference between v0.2 training and v0.5 training;
4. A minimal training script or adapter/fine-tuning recipe;
5. Any notes about the required dataset preprocessing for v0.5.

Even partial information would be extremely helpful for researchers attempting reproduction.

In particular, I am interested in understanding:
- How the multi-view fusion module is trained;
- Whether the TRELLIS.2 backbone is frozen or jointly optimized;
- Whether v0.5 requires full retraining of SS/SLat DiT;
- The amount of data and GPU resources required.

Thank you again for your great work and for making the inference code available.

Best regards.

Start with the v0.5 inference code and compare it with the v0.2 training setup referenced in the issue. The work is complete only when the requested training details, configuration, version differences, preprocessing notes, or a minimal recipe are available, but no specific files or tests are identified.

Written by the indexing model from the issue text.