Wave CLIP Vision Encode

If you're using style models in this pack, you'll need CLIP vision embeddings at some point - and that's where Wave CLIP Vision Encode comes in. It looks like ComfyUI's built-in CLIP Vision Encode node, and it mostly is. The difference is a second output: it stuffs the image embeddings into the wave_config so the wave samplers can select between multiple reference images over time, one per cycle.

How it works

You give it a CLIP vision model, one or more images, a crop mode, and a wave_config. It encodes the image(s) into a standard CLIP_VISION_OUTPUT - exactly what a style model expects - and then stores the full embedding sequence inside a copy of the wave config. If you hand it a batch of reference images, it splits them into a sequence the samplers will cycle through: cycle 0 uses image 0, cycle 1 uses image 1, and so on. That's the trick that lets a morph walk through a style or identity sequence without rebuilding the graph.

Two implementation details worth knowing. It processes images in chunks of 16 to keep VRAM in check, then merges the results back into a single output. And the sequence wraps, so if your batch is shorter than your cycle count it just loops back to the start.

Inputs

• clip_vision - a loaded CLIP vision model (via the standard CLIP Vision Loader).
• image - one image, or a batch for sequence cycling.
• crop - center (default) or none. Center-crop is the standard for CLIP vision preprocessing; choose none if your images are already the aspect ratio you want.
• wave_config - the ongoing wave config, so this node can thread its data back through the same chain.

Outputs: clip_vision_output (wire it into Wave Style Model Apply or any standard style-model node) and wave_config (keep threading it downstream to the sampler).

Where it sits

The README's style path is: Wave CLIP Vision Encode → Wave Style Model Apply → Video/Image Iterative Sampler. The clip vision output feeds the style model; the enriched wave config feeds the sampler so the style strength can wave-modulate per frame. If you never use style models or IPAdapter, you probably don't need this node - the samplers don't require it.

Installing it

Part of kentskooking-nodes: ComfyUI Manager → search kentskooking-nodes → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Kentskooking/kentskooking-nodes

Restart after. No models to download - you bring your own CLIP vision model.

Gotchas

• It requires a wave_config input. Unlike the built-in encode node, you can't just drop it in anywhere - it needs a controller upstream. Forget it and the node errors immediately.
• Chained nodes must keep passing the enriched config. If you branch the wave config off to a preview node and lose the enriched copy, the sampler never sees the sequence. Keep one continuous wave_config line from controller to sampler.
• The crop choice affects what the style model "sees." If your reference has awkward framing, center crop is usually the fix, not the enemy.