Nodes/ComfyUI_IPAdapter_plus_V2/IPAdapter Precise Style Transfer V2
ComfyUI Node

IPAdapter Precise Style Transfer V2

Steal the brushwork, leave the subject alone

By chflame163·Created 2 years ago·Updated 2 years ago· 55
IPAdapter Precise Style Transfer V2
  • model
  • ipadapter
  • image
  • image_negative
  • attn_mask
  • clip_vision
  • MODEL
weight1.00
style_boost1.00
combine_embeds
start_at0.000
end_at1.000
embeds_scaling

The mirror image of the Precise Composition node: where that one pushes the layout, this one pushes the style. IPAdapterPreciseStyleTransferV2 is the advanced apply node for the times you want an image's texture, brushwork, lighting, and palette - its look - while keeping your subject and your prompt firmly in control. The added knob is style_boost, and its default of 1.0 is the giveaway: style transfer is what IP-Adapter does best, so the boost starts on rather than off.

style_boost runs −5 to 5 and scales the style component of the embedding relative to everything else. Crank it past 1 and the output leans harder into the reference's visual language - painterly strokes, color grading, rendering style - even at the cost of some prompt fidelity. Drop it below 1 and you get a lighter seasoning. Negative values push the model away from the reference style, which is a real use case when a reference's palette is fighting you.

The rest of the surface is the standard advanced apply node: weight for overall influence, combine_embeds for how multiple references merge, start_at/end_at for step range, and embeds_scaling for injection mode. Optional image_negative, attn_mask, and clip_vision round it out. One output, MODEL, into your sampler.

How it's different from plain style transfer

The regular weight_type: "style transfer" option on the basic apply node is a blunt switch - it picks a weight curve that emphasizes the image. This node makes the emphasis a continuous, independent dial, so you can keep weight moderate (prompt stays coherent) while nudging style separately. That separation is the point: the community's standing advice is weight 0.6–0.8 for style transfer, and the reason it's a range rather than a setting is that "how much style" and "how much reference" are two different questions. This node lets you answer them separately.

When you'd reach for it

  • Turning a photoreal render into a painterly or anime-styled version of the same composition
  • Matching a brand's or an artist's color grade and rendering across a series
  • Style bleaching - negative style_boost when a reference's look is contaminating output you wanted generic

The honest warning from the KB: IP-Adapter running across all steps softens detail, and heavy style transfer is the worst offender. Keep end_at below 1.0 (say 0.7–0.8) when fine detail matters, or pair it with a sharpening pass downstream.

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

The two knobs interact confusingly. High weight + high style_boost will happily erase your prompt - if the output stops listening to what you typed, drop weight first, then style_boost. And like its composition sibling, this is a single-reference node; for a batch or sequence use IPAdapterPreciseStyleTransferBatchV2, which handles each image separately instead of averaging them into mush.

Categoryipadapter

Inputs (12)

NameTypeDefaultDescription
modelMODEL
ipadapterIPADAPTER
imageIMAGE
weightFLOAT1.00-1–5
style_boostFLOAT1.00-5–5
combine_embedsCOMBO5 options: concat, add, subtract, average, norm average
start_atFLOAT0.0000–1
end_atFLOAT1.0000–1
embeds_scalingCOMBO4 options: V only, K+V, K+V w/ C penalty, K+mean(V) w/ C penalty
image_negativeoptIMAGE
attn_maskoptMASK
clip_visionoptCLIP_VISION

Outputs (1)

NameTypeDescription
MODELMODEL