ComfyUICharacterComposer

ComfyUICharacterComposer isn't a generator and it doesn't call any API. It's a prompt-composition layer that sits in front of Qwen-Image-Edit and turns "rewrite the whole sentence and hope" into a set of dropdowns. You pick a hair color, an outfit, a vibe, and it assembles the final positive and negative prompt strings for you - deterministically, from a seed, with a JSON tag library doing the heavy lifting.

Why bother? Qwen-Image-Edit is where consistent character work actually happens these days - you generate once, then edit that image into new poses and scenes (the whole IP-Adapter stack mostly stopped being the answer). The catch is that its prompt is where people fight it. This node exists for iteration: flip hair_color from blonde to brunette, hit run with the same seed, and everything else stays put. The author pitches it as a workflow layer on top of the Qwen pipeline rather than a replacement, and honestly that's the right framing. The one skeptical comment on their release thread said Qwen AIO is already solid on its own - fair. This buys you speed and reproducibility, not magic.

How it works

Under the hood it's a pure-Python engine with zero dependencies (no requirements.txt, no torch calls). Every run does roughly this:

1. Reads a JSON tag library - tags.json by default, with tags_NSFW.json as the selectable spicier alternative. The tag_file dropdown lists whatever JSON files it finds.
2. Scans your input_prompt for known trait keywords (using the alias map, so blond still matches blonde).
3. Merges that with your UI dropdowns. Every trait input is auto / random / none / a specific value, and random picks are driven by a random.Random(seed) - which is why the same seed reproduces the same character.
4. Applies composition bias and cleanup: composition_mode and smart_preset bias defaults, _composer_rules in the JSON handle conflicts (a kokoshnik headdress drops the ushanka fur hat), and a complexity guard trims low-priority traits when the prompt gets overcrowded.
5. Replaces wildcards - {hair_color} in your base prompt gets substituted with the final choice.

The node is OUTPUT_NODE-style: it prints the assembled prompt to the read-only final_prompt_preview field so you can see what it actually sent, which is the best debug tool this thing has.

The inputs that matter

You don't touch most of the ~40 inputs. Start with these:

• input_prompt - your natural-language base, with {tags} for explicit wildcard injection.
• seed - the reproducibility switch. Same seed, same picks.
• smart_preset / composition_mode - high-level bias. safe portrait, winter fashion, paired cinematic, etc. Leave everything else on auto until you need an exact override.
• extra_modifiers - style flavoring; the default is heavily "gravure / tropical glamour", so expect to edit this if that's not your niche.

Outputs are exactly five: positive_prompt and negative_prompt strings, plus image1 / image2 / image3 passthroughs. Wire the strings into your Qwen-Image-Edit conditioning, and the images into the multi-reference encoder (v2.0 accepts up to three reference images for image-to-image). Leave them disconnected and it works as text-to-image. The bypass_generator toggle returns your raw input_prompt untouched - handy for isolating whether the node or the model is the problem.

Install

Through ComfyUI Manager (search "comfyui-character-composer"), or:

cd ComfyUI/custom_nodes
git clone https://github.com/unh1nge/comfyui-character-composer

Then restart ComfyUI - it appears under Prompt. No pip packages, no model files inside the pack. The catch is the recommended stack: the author targets Qwen-Image-Edit-Rapid-AIO (Phr00t's all-in-one checkpoint) and ships a ready-made QWEN AIO CC 3X WORKFLOW.json. That checkpoint is the real download, and it's a 20B-class model - in practice people run community GGUF quantizations, not bf16, unless you've got the VRAM.

Common gotchas

• The built-in negative prompt is niche-specific. It ships with terms like "no twins, no deformities, no futanari" baked in - written for the author's "cute & tropical gravure" corner. If you're generating something else, don't blindly trust negative_prompt; trim it.
• Random output with no seed discipline. auto traits with fill_auto_traits off leave gaps in your prompt. If a character comes back missing hair, that's usually an unfulfilled auto trait, not the model.
• Qwen drift is not this node's bug. Faces sliding across an edit chain is a Qwen-Image-Edit structural limitation (it re-emits the whole frame). This node can't fix identity drift; it just makes the prompt part reproducible.

It's a niche, opinionated tool - one author's answer to a specific workflow. If you live in Qwen character generation, it's genuinely the fastest way to iterate. If you're on SDXL, it's not for you yet.