Hi, thanks for releasing Cosmos3.
I am trying to reproduce the Cosmos3 PAIBench-C transfer-generation results reported in the technical report/model documentation, especially the single-control segmentation setting.
The report describes the transfer-generation setup as using 50 denoising steps, text guidance 3, control guidance 1.5, shift 10, and full-range CFG. In the released cosmos-framework transfer inference path, I can find the standard guidance / shift / step settings, but I cannot find how to set the separate control-guidance weight described by the two-weight CFG setup.
I opened a related PAIBench issue for the evaluator/reference segmentation artifact side:
SHI-Labs/physical-ai-bench#7
This issue is about the Cosmos3 inference/model side. Could you clarify:
- the exact released-code command/config used for the PAIBench-C Table 16 transfer-generation evaluation
- whether the two-weight CFG / separate control-guidance mechanism is implemented in the public repo
- if implemented, which CLI/config fields correspond to text guidance and control guidance
- the expected frame count and resolution for PAIBench-C reproduction
- whether the reported Cosmos3 Nano/Super PAIBench-C numbers are reproducible from the current public checkpoints and code
- any additional inference settings needed for the segmentation-control run
The goal is to distinguish a released-code Cosmos3 PAIBench-C reproduction from an approximate run using the public transfer inference defaults.
Thanks.
Hi, thanks for releasing Cosmos3.
I am trying to reproduce the Cosmos3 PAIBench-C transfer-generation results reported in the technical report/model documentation, especially the single-control segmentation setting.
The report describes the transfer-generation setup as using 50 denoising steps, text guidance 3, control guidance 1.5, shift 10, and full-range CFG. In the released
cosmos-frameworktransfer inference path, I can find the standard guidance / shift / step settings, but I cannot find how to set the separate control-guidance weight described by the two-weight CFG setup.I opened a related PAIBench issue for the evaluator/reference segmentation artifact side:
SHI-Labs/physical-ai-bench#7
This issue is about the Cosmos3 inference/model side. Could you clarify:
The goal is to distinguish a released-code Cosmos3 PAIBench-C reproduction from an approximate run using the public transfer inference defaults.
Thanks.