IPAdapter Precise Style Transfer Batch V2

The batch counterpart to IPAdapterPreciseStyleTransferV2: same style_boost dial for pushing the reference's visual language - texture, brushwork, palette - but built for a batch of reference images, each conditioned on its own frame. Where the single version has one image input and one style to transfer, this one treats every image in the image batch as its own style job, which is the difference between "stylize this sequence consistently" and "stylize each frame from its own source." Under the hood that's unfold_batch = True - the standard batch flag this pack uses to keep per-image conditioning separate instead of averaging the batch into one blob.

Inputs match the single version: image, weight, style_boost (−5 to 5, default 1.0 - note the contrast with the composition node, which defaults to 0; style transfer is the default emphasis here), combine_embeds, start_at, end_at, embeds_scaling, plus optional image_negative, attn_mask, clip_vision. Output is MODEL. Everything in the single-image article transfers: positive style_boost leans harder into the reference's look at some prompt-fidelity cost, negative pushes the style away, and small steps are the way to move it.

When you'd reach for it

The realistic job is animation with per-frame style references - a keyframe sequence where each frame's palette and texture come from a matching style image - or batch variation generation where each variant pairs with its own style reference. If you only have one style image, the single version is the right tool and this one buys you nothing.

Install

Part of chflame163/ComfyUI_IPAdapter_plus_V2, the V2 fork of cubiq's pack. Manager → search ComfyUI_IPAdapter_plus_V2, or:

cd ComfyUI/custom_nodes
git clone https://github.com/chflame163/ComfyUI_IPAdapter_plus_V2

Restart. No pip requirements.

Gotchas

Same batch-sync rule as every batch node here: the reference batch must align with your frame count, or frames quietly borrow each other's styles. And the style-overdose failure is easier to hit in batch - if every frame comes out looking like the same smeared painting, you've got style_boost and weight both too high and the model has stopped listening to the prompt. Drop weight first. Finally, respect end_at: heavy style transfer across all sampling steps is the biggest detail-killer in the IP-Adapter family, and in a sequence it compounds frame over frame.