HY OmniWeaving Image Prep

HY OmniWeaving Image Prep is the least glamorous node in this pack and one of the most important. It takes your reference image and resizes it to the workflow's canvas using Lanczos interpolation and a center crop - the exact preprocessing the original OmniWeaving repo applies. That's it. One image in, one image out.

Why does a resize node deserve your attention? Because OmniWeaving is picky about parity. The validated community workflows were tuned against the official repo's pixel pipeline, and the whole pack's design philosophy is "keep sampling/CFG/VRAM on stock ComfyUI, reimplement only OmniWeaving-specific semantics." Image preprocessing is one of those semantics. If you feed a reference image at a random resolution and let the VAE or CLIP-vision encoder decide how to handle it, your first frame can silently drift from the reference - the kind of bug that reads as "the model isn't following my image" when the real culprit is a resize mismatch. This node exists to take that variable off the table.

What it does, precisely

_prepare_omniweaving_images runs common_upscale(..., "lanczos", "center") to your target width and height - the same call the Conditioning node uses internally on its own references. So it's the blessed image entry point for any i2v or reference2v-style run: feed it your reference, and its prepared_images output goes into HY OmniWeaving Redux Vision Encode (for the CLIP-vision path) or HY OmniWeaving Conditioning (as reference_images).

Inputs: reference_images, width, height (defaults 848x480, step 16, matching the pack's standard canvas). If you're using the pack's own workflows, leave the defaults alone.

Installing it

It ships in hy_omniweaving_comfyui_unofficial. 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; needs a recent ComfyUI for the extension API. It doesn't add any model downloads of its own - it's pure tensor math - but the workflow it feeds still needs the OmniWeaving diffusion model, text encoders, vision pair, and VAE from the README.

When you can skip it

If your reference is already exactly the workflow resolution and you don't care about perfect parity with the official repo, the Conditioning node will re-prep it anyway - you can skip the explicit prep node and wire your image straight in. But the pack's README calls it "the blessed image-entry point for parity-sensitive i2v / reference2v style runs," and honestly it's cheap insurance. When video models act up for no visible reason, "the input wasn't preprocessed the way the author's reference run preprocessed it" is a very common answer. Use it.

The one trap: don't confuse this with HY OmniWeaving I2V Semantic Images. Prep just resizes and crops - it stays in pixel space. The semantic node is the one that roundtrips through the VAE to build the model's understanding of the first frame. Prep first, semantic second, in the blessed order.