Smart Style Model Apply DUAL (CRT)

Prompting gets you a subject; style models get you a look. Smart Style Model Apply DUAL (CRT) applies an image-based style to your conditioning - think "make this prompt render in the style of this reference painting" - and it does it for one or two reference images at once, each with its own strength. Feed it a style model, a CLIP Vision model, and a reference, and the conditioning that comes out carries that style into the sampler.

It's the CRT-Nodes take on ComfyUI's built-in StyleModelApply (the UnCLIP-style conditioning path), with the CLIP Vision encoding handled for you and a second reference image layered in. The "DUAL" in the name is the whole pitch: two style references, independently weighted, blended into one conditioning.

How it works

At its core it does what StyleModelApply does - but in one node with convenience built in:

1. Encode the reference image(s) with the provided clip_vision model (crop set to center trims to a square for encoding; none leaves the full frame).
2. Run the style_model over the encoded image to produce style conditioning.
3. Apply it to your text conditioning at strength_1 (and strength_2 for the second image), using one of two strength modes:
   • multiply - scales the style conditioning directly by the strength (the blunt, classic approach).
   • attn_bias - injects the strength as an attention bias on the conditioning's attention mask. This is the ComfyUI-native path for style models and is generally the better behaved of the two.

The second image is optional: connect image_2 and it activates automatically; leave it empty and it's effectively a single-style apply. The node also caches only its most recent result, so it stays light.

Inputs and output

• conditioning - your text conditioning from a CLIP text encode.
• style_model - a STYLE_MODEL (the same kind ComfyUI's StyleModelApply loads).
• clip_vision - a CLIP_VISION model for encoding the reference.
• image_1 (required) and image_2 (optional) - the reference images.
• strength_1 / strength_2 (0–10, default 1) - per-reference influence.
• strength_type - multiply or attn_bias.
• crop - center or none for the CLIP Vision encode.

Output is a single conditioning, ready to feed your KSampler alongside (or instead of) your plain positive conditioning.

Where you'll use it

• Style transfer: "same subject, painterly like this reference."
• Character/style consistency across a batch - one fixed reference, one fixed seed, a folder of prompts.
• Blending two aesthetics: image_1 at 0.7, image_2 at 0.4, and you're mixing looks with dials instead of Photoshop.

Gotchas

• It needs a style model and a CLIP Vision model loaded in - this is a conditional-style node, not a magic wand. If the menu is missing it entirely, you probably don't have the model types wired up, not the node.
• multiply vs attn_bias genuinely changes the result; attn_bias is the better default with ComfyUI's style models, so reach for multiply only if you've compared and it works for your base.
• Strength is not linear in the way you'd hope - halving 1.0 to 0.5 doesn't necessarily half the stylistic effect. Tune by eye.

Install is pack-level: ComfyUI Manager → search CRT-Nodes → install and restart, or clone the repo into custom_nodes. Heavy suite, graceful degradation on missing optional deps.