HY OmniWeaving UNet Loader

If you try to load an HY-OmniWeaving diffusion checkpoint with the stock UNETLoader, you get an ugly pile of unexpected key warnings and - depending on the file - a model that silently doesn't work. That's because Tencent's OmniWeaving checkpoints store attention weights as separate img_attn_q/k/v and txt_attn_q/k/v tensors, while stock ComfyUI's HunyuanVideo 1.5 path expects them fused as qkv. This node exists to bridge that gap.

HY-OmniWeaving is Tencent's April 2026 reason-then-generate video model: a fine-tune of HunyuanVideo 1.5 with a Qwen2.5-VL MLLM in front that parses your prompt and reference frames, then forwards hidden states into the diffusion transformer through an mm_in MLP connector. The official pipeline wants eight GPUs. This pack is the community's ComfyUI extraction lane - it keeps sampling, CFG, and VRAM management on stock nodes and reimplements only the OmniWeaving-specific bits. The UNet loader is one of those bits.

How it works

Under the hood, HY OmniWeaving UNet Loader does three jobs:

1. Attention-key fusion. It scans the state dict for split img_attn_q/k/v and txt_attn_q/k/v tensors inside each double_blocks layer and torch.cats them into the qkv layout ComfyUI's transformer expects. With strict_mode on (the default), a partially-fused checkpoint raises an error instead of loading a half-broken model - flip it off only if you know what you're doing and want it to limp along.
2. mm_in extraction and attachment. OmniWeaving's mm_in connector weights are present in the checkpoint but stock ComfyUI ignores them (that's what the unet unexpected: ['mm_in...'] diagnostic warning means). The loader pulls them out, attaches them to the model, and registers the deepstack-style wrapper that lets the diffusion model accept projected hidden states.
3. Weight dtype control. weight_dtype defaults to default, but you can force fp8_e4m3fn or fp8_e4m3fn_fast (adds the fp8-optimization path) or fp8_e5m2. The README's blessed checkpoint is already the fp8_e4m3fn_patched build, so most people leave this alone.

The inputs that matter

Only two things actually matter for a beginner:

• unet_name - pick the OmniWeaving transformer from your models/diffusion_models folder. The pack assumes it's named hy_omniweaving_hunyuanvideo15_transformer_fp8_e4m3fn_patched.safetensors.
• strict_mode - leave it on.

The single MODEL output wires straight into your KSampler, exactly like a stock UNet loader's.

Installing it

This node ships in the hy_omniweaving_comfyui_unofficial pack. ComfyUI Manager: search "hy_omniweaving" and install, or clone it manually:

cd ComfyUI/custom_nodes
git clone https://github.com/Shiba-2-shiba/hy_omniweaving_comfyui_unofficial

Then restart ComfyUI. The pack has no extra pip dependencies - requirements.txt is effectively empty - but it does need a recent ComfyUI checkout because it registers through the newer comfy_entrypoint extension API, not the old NODE_CLASS_MAPPINGS mechanism. The model files come from HuggingFace; the exact download mapping is in the pack README (Shiba-2-shiba's FP8-patched repo, plus the tencent/HY-OmniWeaving and vafipas663/HY-OmniWeaving_repackaged repos).

Common issues

• "unet unexpected: ['mm_in...']" in the console - that's expected and harmless; it's the stock loader saying it doesn't know mm_in.*. That's precisely why this node exists. Don't panic, don't go hunting for another loader.
• "Failed to load HY-OmniWeaving diffusion model" - you pointed it at a non-OmniWeaving Hunyuan 1.5 file, or the checkpoint name differs from the validated workflow's. Use the exact filename in the README.
• Strict mode errors about partial attention tensors - the file is a half-fused build; grab the FP8-patched checkpoint the README links.

One honest heads-up from the community: OmniWeaving is slow even on an RTX 6000, and wants CFG plus 30–50 steps to look right. The loader is the easy part. Give yourself time on the sampling side.