Composition (Single)

Composition is the "how the frame is arranged" dimension - rule of thirds, golden ratio, symmetry, centered, frame-within-frame, leading lines, vanishing point. Twenty classic framing techniques, each appended to your base prompt by ID. It's a quiet node in a pack of louder ones, but composition words are some of the oldest and most reliable vocabulary in image generation, so it punches above its apparent weight.

The modifiers read like a photography textbook index: 1 is "rule of thirds, balanced placement, classic composition," 2 is "golden ratio, Fibonacci spiral, mathematical beauty," 3 is "symmetrical composition, mirror balance, formal structure," 4 is "centered composition, bull's eye framing, direct focus," 5 is "frame within frame, layered composition," 6 is diagonal lines, 7 is "leading lines, eye guidance, directional flow," 8 is "vanishing point, perspective depth, convergence." Each is a compact phrase built to slot into a comma-joined prompt. Note the default here is 1 (rule of thirds), which is the one genuinely safe default in the whole pack - most of the others assume a subject that benefits from drama.

The inputs: text (base prompt), modifier_id (1–20, default 1), and preserve_base (default true). Output: single STRING, modified_prompt.

Install is the pack-standard zero-dependency story: ComfyUI Manager → search "Variationator" → install → restart, or cd ComfyUI/custom_nodes && git clone https://github.com/MushroomFleet/ComfyUI-Variationator.git. No models; the install-portable.bat references a requirements.txt that isn't shipped, so there's nothing to pip-install. Look under Prompt Orchestration.

Here's the honest engineering reality, and it's the same for every prompt-level composition approach: words can suggest a composition, but they can't enforce one. "Rule of thirds" as a phrase nudge works best on checkpoints that saw a lot of photography text; on a model that doesn't map framing vocabulary, you'll get the same layout regardless of ID, and the fix is not more composition words - it's a ControlNet with a pose or depth map if you need real framing guarantees. The KB's ControlNet coverage is the right read if you hit that wall. Where this node does shine is batch A/B work: run one subject through rule of thirds vs. symmetry vs. centered and see which framing vocabulary your checkpoint actually respects, then build around the winners. That's a genuinely useful five-minute experiment.