HY OmniWeaving Redux Vision Encode

Reference-to-video is where OmniWeaving shines, and reference images have to get into the model somehow. The official pipeline routes them through a CLIP-vision encoder that projects image features into the text space. HY OmniWeaving Redux Vision Encode is this pack's local, fully-offline version of that route: it runs a SigLIP image encoder plus a Redux-style image embedder - the same Flex-Redux pairing, from the Runware/Flex-Redux HF repo - and emits a single CLIP_VISION_OUTPUT with mm_projected populated. No API, no cloud, no hidden server. It's all weights on your disk.

What that output actually is: the SigLIP encoder turns each image into a sequence of vision features (last_hidden_state, plus penultimate_hidden_states and all_hidden_states), and the Redux embedder squashes those into mm_projected - the projection that the Qwen text encoder can consume as if it were text tokens. That's the deepstack-friendly path the whole pack is built around: mm_projected being None is the classic symptom of a legacy or fallback CLIP-vision route, and this node exists to keep that from happening.

The inputs that matter

• images - the reference image (or batch of them). Wire prepared_images from HY OmniWeaving Image Prep if you want pixel-exact parity; that node does the same Lanczos + center-crop the original repo does.
• image_encoder_model and image_embedder_model - two dropdowns over your models/clip_vision folder. The pack tries to auto-pick image_encoder.safetensors and image_embedder.safetensors (the README filenames from Flex-Redux). Both must be present - the encoder alone is useless without the embedder.
• crop - center (default) or none. Keep center; it's the parity default.
• device - default or cpu, advanced, for forced offload debugging.

Output is a single CLIP_VISION_OUTPUT. It wires into HY OmniWeaving Conditioning (which forwards it into the conditioning for non-t2v tasks) or into HY OmniWeaving Clip Vision Concat when you're stacking multiple references.

Installing it

In the hy_omniweaving_comfyui_unofficial pack. ComfyUI Manager (search "hy_omniweaving") or:

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

Restart. No pip deps - the SigLIP model is loaded through transformers, which any modern ComfyUI install already has. You do need a recent ComfyUI for the extension API, and the two clip_vision files from Flex-Redux in the README.

Common issues

• "Failed to load Redux image encoder cleanly" - you selected the wrong file for one of the two dropdowns (e.g. an embedder where it wants the encoder, or a random CLIP file). The encoder is SigLIP, the embedder is the diffusion_pytorch_model.safetensors Redux weights.
• Only one of the two files downloaded - the node will fail at the missing one. Grab both.
• mm_projected still None downstream - that's a signal you're hitting a fallback CLIP-vision route, not this node. If you connected this node and still see it, check the wire into Conditioning; for t2v it's intentionally ignored.

One practical note from the pack's own docs: in the current validated i2v path, text-side multimodal input prefers explicit semantic_images even when Redux clip_vision_output is also present. So this node is not always the star of the show - but for reference2v and multi-reference runs, it's exactly the piece the official behavior maps onto.