Style Transform

People hear "Style Transform" and expect an actual model that re-renders their image in a new style. It's not that - it's a text transformer, and that's fine, because it costs you zero VRAM and no extra model. You feed it a prompt, pick from ten baked-in artistic styles (or type your own), and it rewrites the prompt to push the sampler in that direction. Same subject, new aesthetic, all in the words.

The mechanism is a template lookup plus string assembly. The pack ships a dict of style descriptions - cyberpunk maps to "cyberpunk aesthetic with neon lights, high-tech elements, and dystopian atmosphere", noir to "film noir style with dramatic shadows, high contrast, and moody atmosphere", impressionist, bauhaus, art_deco, brutalist, minimalist, futuristic, vintage, pop_art_1960s - and then the output depends on one toggle:

• maintain_composition=True produces an instruction-style sentence: Transform to {style} while maintaining the original composition. Base scene: {your prompt}. This one is genuinely suited to the newer LLM-encoded checkpoints, which read prompts as instructions rather than token bags.
• maintain_composition=False goes back to classic comma joining: {your prompt}, {style description} - the safer bet on tag-based SD 1.5/SDXL models.

Select custom in the style dropdown and the optional custom_style field opens up, letting you describe literally any look in your own words instead of using the ten presets.

The inputs that matter: text (your base prompt), style (dropdown of ten presets plus custom), maintain_composition (the true/false split above), and custom_style (only read when style is custom and the field is non-empty). Output is a single STRING, styled_prompt, that feeds any prompt input in the graph. The README's example is a nice sanity check: "Vintage motorcycle" through cyberpunk should come out as a neon-and-tech variant of the same scene.

Installing is the whole-pack story - it's a zero-dependency custom node, pure Python, no models to download. ComfyUI Manager: search "Variationator", install, restart. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/MushroomFleet/ComfyUI-Variationator.git. There's an install-portable.bat that tries pip install -r requirements.txt, but the repo has no requirements.txt - ignore that step, it isn't needed. You'll find the node under Prompt Orchestration.

Two honest takes. First: because this node writes whole instruction sentences when maintain_composition is on, it plays nicer with modern LLM-encoded models than the comma-soup modifier nodes do - that's the one place in this pack where the "old vs new prompt style" divide actually flips. Second, don't expect a miracle on a checkpoint that never saw "impressionist painting" in training. Style-in-prompt is a nudge, not a guarantee; if the result isn't pushing far enough, that's a prompt-weight problem, not a node bug. And like everything in Variationator, over-stacking styles alongside a dozen modifiers just ends in prompt soup - one style transform plus a couple of modifiers is usually the ceiling before things blur together.