HY OmniWeaving I2V Semantic Images

OmniWeaving's image-to-video path has a quirk you don't see in most I2V models: the first frame the text side of the model "sees" isn't your raw reference image. It's a semantic image - a reconstructed version of your reference, derived by roundtripping it through the VAE. HY OmniWeaving I2V Semantic Images is the node that produces that reconstruction, so the same semantic frame can feed both the text-side multimodal input and the CLIP-vision encoding. If that sentence made your eyes glaze over, here's the practical version: it's the node that keeps the model's mental picture of your first frame in sync, which is a big chunk of why OmniWeaving follows references as well as it does.

How it works

The roundtrip is exactly what the name suggests. The reference image gets Lanczos-resized and center-cropped to your width/height (the same prep as HY OmniWeaving Image Prep), VAE-encoded into latent space, then immediately VAE-decoded back to pixels. That encode→decode trip is a lossy reconstruct - it keeps the semantic content (subject, layout, composition) while smoothing out the kind of pixel-level detail the VAE doesn't care about. The result is the semantic image the OmniWeaving pipeline expects.

Then it gets one more trick: that decoded frame is re-encoded into a single-frame latent, which is what actually anchors the output video's first frame. So you get a semantic_images output (an IMAGE) that wires into HY OmniWeaving Text Encode's semantic_images input - the text-side multimodal route the current validated i2v workflow prefers - while the latent side is handled separately by HY OmniWeaving Conditioning.

Inputs are minimal: vae (from HY OmniWeaving VAE Loader), reference_images, width, height (848x480 default). One IMAGE output.

Installing it

Same pack as everything here. ComfyUI Manager (search "hy_omniweaving") or:

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

Restart after. No pip dependencies, recent ComfyUI required for the extension API. The only model dependency is the VAE it needs as input - the hunyuanvideo15_vae_fp32.safetensors from the README.

Common issues

• "Expected video latent tensor" or a rank error - the VAE you wired in isn't producing the 4D/5D latent this node expects. Use the pack's VAE loader, not an arbitrary one.
• Semantic image looks blurry / washed out - that's not a bug, it's the roundtrip working as intended. The reconstruction is supposed to be a smoothed semantic version of your frame, not a photoreal copy. If you want pixel fidelity in the actual output, that's what the latent anchoring in Conditioning handles.
• Confusing it with Image Prep - prep resizes in pixel space and stops there. This node resizes and roundtrips through the VAE. Both belong in the blessed i2v order: prep first (or feed raw), then this, then Redux Vision Encode, then Text Encode.

One honest note: for simple i2v the roundtrip mostly matters for parity with the official repo's behavior. The pack's own docs say text-side multimodal input prefers explicit semantic_images when connected, so if a workflow's output drifts, the usual first suspects are task mismatch between Text Encode and Conditioning - not this node.