Stability Control Style

Control Style is the reference-image node of the pack: feed it a picture whose look you want - a painting's brushwork, a film still's grading, a render's material feel - plus a prompt for content, and it generates a new scene in that style. It's the sibling of Control Structure (which preserves layout) and Control Sketch (which preserves drawn lines), but this one's after the surface: texture, color, mood, finish.

It's a cloud node from the official "Stability API nodes for ComfyUI" pack, calling stable-image/control/style on Stability's servers - zero VRAM, nothing downloaded, credits per call. The local analog is a reference-style ControlNet or a style-transfer pass, which gives you more knobs but also more setup and a real VRAM bill. The community's default posture toward this pack is "not local, not interested," and it's a fair position if you have the hardware. This node is for the no-hardware, no-setup route.

How it works

The reference image supplies the style; your prompt supplies the subject and scene. The server blends them. One thing to know about this node specifically: it's the exception in the pack where the input image doesn't switch the request into strict image-to-image mode - here the reference is purely a style source, and the output dimensions follow your chosen aspect_ratio, not the input image's.

The inputs that matter:

• image - your style reference. Quality matters more than content; a clean, single-style image transfers far better than a cluttered one.
• prompt - the new content. This is where you say what the scene actually is.
• fidelity (0.0–1.0, default 0.5) - how closely the output matches the reference style. Low means the style is a loose flavor; high means the output really commits to it.
• aspect_ratio - the usual nine presets (1:1, 16:9, etc.), which is a nice freedom you don't get with most of the pack's edit nodes.
• negative_prompt, seed - the standard pair.

Plus the shared style/style_preset (ignored unless style is on) and output_format. Output is a single IMAGE tensor.

Installing it

From the "Stability API nodes for ComfyUI" pack - ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Stability-AI/ComfyUI-SAI_API
# restart ComfyUI

No models download; requests is the only real dependency beyond ComfyUI's stack. Key via SAI_API_KEY, a sai_platform_key.txt file, or the per-node api_key_override (don't share workflows using that - the key rides in the JSON).

Common issues

• Style barely shows - fidelity too low, or the prompt is fighting the style with heavy content descriptions. Drop the prompt to something plain and raise fidelity.
• Output looks like a photo with a filter - you're describing the scene too literally. The style needs room; give the prompt less to do.
• Wrong node - keeping the exact composition of your reference (a person's pose, an object's placement) is Control Structure's job, not this one's. This node keeps the look and changes everything else.

Style transfer was never the most reliable trick in the box, and a hosted version trades away your local fine-tuning for convenience. But for "make this prompt look like that painting," it's one node and one slider - and that's a genuinely hard thing to set up locally for the same effort.