Community extraction from NitroFusion2, revision 62d1a7000af9c46a0c4348d936bd7670ff33c1b9. The 15 stored tensors preserve the original names, BF16 values and shapes of all_x_embedder.*, all_final_layer.*, t_embedder.*, cap_embedder.*, x_pad_token, and cap_pad_token. See tensor_mapping.json and extraction.json for the exact source ranges and digest.
These adapters map 2×2 image latent patches to transformer embeddings, map caption and timestep features into conditioning embeddings, provide learned padding, and map shared-transformer hidden states back to latent patch predictions. They are not a standalone image generator or a complete DiT. The 30-layer trunk, the refiner stacks, VAE, text encoder and scheduler are excluded. Use the original pipeline for patch ordering, timestep scaling, positional encoding, masks, padding and output reconstruction.
Install PyTorch, accelerate, safetensors and diffusers==0.36.0; component.load_model(directory) returns the source-compatible adapters and loads all tensors strictly. The checkpoint was tested in BF16 on CPU: a COCO image was encoded with the source-compatible VAE, then passed through the latent patch embedder; timestep and caption adapters were also checked, and the final output layer was exercised with synthetic transformer hidden states. The test passed with finite outputs and deterministic repeats. See validation.json for shapes and details. Synthetic hidden states do not validate full model generation or parity.
Apache-2.0; see the included LICENSE and NOTICE files. Preserve upstream attribution and applicable third-party terms.
ChenDY/NitroFusion2