Cosmos-Predict2¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
cosmos-predict2-2b-t2i |
nvidia/Cosmos-Predict2-2B-Text2Image | acdb5fd |
NVIDIA Open Model License | Verified |
Catalog: cosmos-predict2-2b-t2i ·
Models: cosmos-predict2-2b-t2i · Task: t2i
Architecture. Cosmos-Predict2 combines a 2B, 28-layer 3D diffusion transformer with a T5-XXL text encoder, a WAN causal VAE, and Karras flow-match Euler sampling. The published transformer and VAE are float32; the two T5 shards are bfloat16 and are upcast for CPU float32 inference.
Usage. Accept the gated repository terms, then run
python -m imagehub.pretrained cosmos-predict2-2b-t2i checkpoints/cosmos-predict2-2b-t2i.
Settings. steps (35), guidance (7.0; 1 and below disables classifier-free guidance),
width (1360), height (768), negative_prompt, max_sequence_length (512), seed, and
latents. Width and height must be multiples of 16.
Validation. Published headers match all 195 T5, 567 transformer, and 194 VAE tensors. Exact random-weight tests cover each component, the tokenizer, Karras schedule, guidance on and off, non-square output, two seeds, and whole-pipeline uint8 output. Real-checkpoint component and end-to-end tests are provided for the validation run.
Limits and license. The backend supports one prompt and one image per call. Batches, image conditioning, video, training, callbacks, and safety-checker integration are out of scope. The gated weights use the NVIDIA Open Model License; review its safety and attribution conditions before use.