Post training

[tarcisio16]

I’m interested in post-training existing Diffusion Renderer checkpoints to potentially improve visual realism, generate new g-buffers, and condition the diffusion process with additional inputs (e.g., canny edge maps) using cosmos transfer.

•	Can the current training script be used directly for this type of fine-tuning, or are there any specific caveats I should be aware of?
•	Are there plans to release more detailed training documentation or implementation notes specific to the Diffusion Renderer setup?

[jayhsu0627]

I assume they used the cosmos-predict 1 post-training code.