Chroma Style Model

Same disclaimer as this pack's other Chroma-prefixed nodes: the README marks the whole Chroma section of ComfyUI_FluxMod as deprecated in favor of ComfyUI's native Chroma workflow. Worth saying up front so you're not chasing this node down assuming it's the current recommended path - it's the pack's early, self-contained answer to something ComfyUI core has since absorbed.

What it's answering is style-model conditioning - the same idea behind Flux's official Redux tool: instead of (or alongside) a text prompt, you feed in a reference image and let its visual "essence" steer generation. You run the image through a CLIP vision encoder, run that output through a style model, and this node folds the result into your text conditioning before it hits the sampler. This is a Chroma-specific fork of that mechanism rather than the generic ComfyUI StyleModelApply node, most likely because Chroma's architecture diverges from stock Flux in ways that matter here - it's a heavily reworked DiT, not a plain Flux checkpoint, and other Chroma-adjacent integrations (PuLID being the documented example) needed compatibility patches for exactly that reason.

What it does

You give it three things: your existing conditioning, a style_model (the loaded style/Redux-style model), and a clip_vision_output (the encoded reference image). Two float knobs control how much the style bleeds through. strength (default 1, range -10 to 10) is the overall intensity - negative values push away from the reference rather than toward it, if you want that effect. truncate_percent (default 1, range -1 to 1) is more specific: its tooltip says it "truncates clipvision conditioning to the first truncate_percent values when > 0" and "truncates the last |truncate_percent| values when < 0" - in plain terms, it trims how much of the CLIP vision embedding sequence actually gets used, which is the practical way to dial a style reference down from "basically recreate this image" toward "just take a hint from it." Output is a single CONDITIONING you route into your sampler in place of (or merged with) your plain text conditioning.

How to install it

Via ComfyUI Manager: search "ComfyUI_FluxMod," install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/lodestone-rock/ComfyUI_FluxMod.git

Restart and refresh. You'll also need a compatible style model file and a CLIP vision encoder set up upstream - this node doesn't load either, it just applies them.

Common issues & troubleshooting

Style overwhelms your prompt entirely. This is the classic complaint with any Redux-style workflow, not something specific to Chroma: even a moderate strength can dominate the text conditioning. Pull strength down first, then experiment with truncate_percent to trim how much of the reference actually gets injected - a lower value keeps more of your text prompt intact.

Node errors on style_model or clip_vision_output mismatch. Both of those are typed inputs (STYLE_MODEL, CLIP_VISION_OUTPUT) that have to come from matching upstream nodes - a style model loader and a CLIP vision encode step feeding an image in. If either chain is missing or the model files aren't the pair the workflow expects, you'll get a type error at graph-build time rather than a bad image.

Getting Chroma-appropriate results. Chroma is documented as wanting verbose, descriptive text prompting more than stock Flux does - that habit doesn't go away just because you're also feeding in a style reference. Keep your positive prompt detailed even when leaning on this node; it's meant to steer, not replace, the text side.