| WorldStereo-memory-dmd-fp8_e4m3fn_scaled |
| ======================================== |
|
|
| This work is a derivative of two upstream projects: |
|
|
| 1. hanshanxue/WorldStereo (worldstereo-memory-dmd variant) |
| https://huggingface.co/hanshanxue/WorldStereo |
| License: MIT |
| Copyright (c) hanshanxue and contributors |
|
|
| 2. Wan-AI/Wan2.1-I2V-14B-480P-Diffusers (base model, frozen in WorldStereo) |
| https://huggingface.co/Wan-AI/Wan2.1-I2V-14B-480P-Diffusers |
| License: Apache License, Version 2.0 |
| Copyright (c) Alibaba Group (Wan team) |
|
|
| Statement of Changes (Apache-2.0, Section 4): |
| --------------------------------------------- |
| The model weights in this repository have been quantized from bfloat16 to |
| torch.float8_e4m3fn with per-tensor scaling ("fp8_e4m3fn_scaled"), following |
| the convention defined in kijai/ComfyUI-WanVideoWrapper. No tensor values |
| were intentionally altered beyond the fp8 quantization noise. No layers were |
| added, removed, or restructured. See README.md for the exact quantization |
| recipe (exclusion list, scale formula, marker tensor) and a list of which |
| tensors were cast vs preserved at full precision. |
|
|
| The original tensor data is freely available at the source repositories |
| linked above. |
|
|
