HY-OmniWeaving HunyuanVideo 1.5 FP8 Patched

This repository provides a patched version of:

• hy_omniweaving_hunyuanvideo15_transformer_fp8_e4m3fn.safetensors

The public checkpoint I started from had zeroed mm_in.linear_2.weight and mm_in.linear_2.bias.
This patched file restores those two tensors by transplanting them from an updated base model where that layer is present and non-zero.

Original upstream model:

• https://huggingface.co/tencent/HY-OmniWeaving

What Was Changed

Only the following tensors were replaced in the FP8 checkpoint:

• mm_in.linear_2.weight
• mm_in.linear_2.bias

The purpose of this patch is to restore the missing / inactive mm_in.linear_2 connector so the model no longer carries an all-zero layer at that point.

This is not a retrain or finetune. It is a targeted tensor transplant patch.

Verification

The original FP8 checkpoint showed:

• mm_in.linear_2.weight: zero ratio 1.0, norm 0.0
• mm_in.linear_2.bias: zero ratio 1.0, norm 0.0

The patched checkpoint shows:

• mm_in.linear_2.weight: zero ratio 0.0
• mm_in.linear_2.bias: zero ratio 0.0

The patched tensor stats match the source base model for those two tensors.

Files

• hy_omniweaving_hunyuanvideo15_transformer_fp8_e4m3fn_patched.safetensors: patched checkpoint

Usage

Use this file anywhere you would normally use the original FP8 transformer checkpoint for HY-OmniWeaving / HunyuanVideo 1.5 workflows, including ComfyUI setups that expect this single-file transformer format.

If your workflow currently points to the original:

• hy_omniweaving_hunyuanvideo15_transformer_fp8_e4m3fn.safetensors

replace it with:

• hy_omniweaving_hunyuanvideo15_transformer_fp8_e4m3fn_patched.safetensors

Notes

• This patch only addresses the zeroed mm_in.linear_2 tensors.
• No claim is made here that it fixes every quality or parity issue in HY-OmniWeaving workflows.
• Please follow the original upstream model's license, usage terms, and attribution requirements.