condition image resolution

[Sugercanee]

great work!
what resolution you used for the condition image during training, and was it rendered using TRELLIS's render_cond.py script?

[Mengmouxu]

During the training process, the conditional image will be resized to (518, 518). This setting was implemented in the data loader. Since our final training process did not adopt single-asset mixed training, we could skip using the TRELLIS render_cond.py script.

[Sugercanee]

Thanks for your answer. When training, at what number of scene assets do you run will out of memory?

[Mengmouxu]

We set the maximum number of assets per scene to N′ = 7 on a single A100 GPU

[Sugercanee]

ok, I find it in paper. Have you tried inference tests on scenes with more than 8 assets? Does it work well?

[Mengmouxu]

Larger asset numbers may lead to out-of-memory (OOM) issues. In our experiments, we set a limit of up to 7 assets per scene, as random selection in scenes with more assets is sufficient to cover relevant data, and 7 assets are already enough for our experiments. Of course, using a graphics card with larger VRAM can support training for scenes with more assets.