Character Reference Prepare

Character reference workflows have an annoying branch point early on. Sometimes you already have a clean reference photo and you just want to use it. Sometimes you have a messy one and need a version with the jacket removed and the background gone. Sometimes you have no photo at all and only a description. Character Reference Prepare [Eclipse] is the node that makes that choice explicit, per role, instead of you copy-pasting four near-identical subgraphs.

The mode is the node

It has two widgets that matter: role (face, body, clothing, rear) and mode:

• reference - pass the source through untouched. No preparation sampler runs.
• extract - use the source image to generate an isolated reference of that role.
• generate - build the reference from text alone, ignoring any source.
• disabled - omit this role entirely, and generate nothing.

That last one is the quietly useful option. Instead of muting or bypassing a whole branch when a character has no rear reference, you set disabled and Character Reference Select returns no image downstream, leaving the graph readable.

role picks which section of prompts/character_references.json supplies the base positive and the role-specific negative - so extract on clothing asks for a garment-only pass, while face asks for identity, hair and skin tone. The file itself is extracted into ComfyUI_Eclipse/prompts/ on first startup, and edits are read on the next execution rather than at import.

Inputs

description is where you add specifics - facial features, proportions, "remove the print on the shirt". It appends to the file's positive; it doesn't replace it. instructions appends after that, for what to preserve or exclude. Both are ignored and hidden once you connect positive_override.

The image input is optional in the schema but required for reference and extract. It takes exactly one image - no batches, no lists. Crop, resize and preview upstream, because this node does not crop for you.

The two override inputs are the escape hatch. positive_override replaces the entire preparation positive, file text, description and instructions included, and it's passed through exactly as supplied; an empty connected string clears it. negative_override replaces this role's negative independently, and an empty string means no negative. Both are plain STRING sockets meant for String Multiline nodes, and both live in your workflow, not the file - editing the JSON will never update a connected override.

Outputs

• reference - an ECLIPSE_CHARACTER_REFERENCE control object. It is not an image and not a prompt. Wire it to Character Reference Select.
• positive - the preparation prompt for the encoder in extract/generate mode.
• negative - the matching negative for that same encoder.
• source - the full source image for the extraction encoder. Present in reference/extract, empty in generate/disabled.

The wiring that works: Prepare's positive/negative → an encoder and sampler for that one role → the decoded result into Select's generated input → Select's image output into the matching socket on Character Reference Pack.

Install

ComfyUI Manager, search ComfyUI_Eclipse, or:

cd ComfyUI/custom_nodes
git clone https://github.com/r-vage/ComfyUI_Eclipse

ComfyUI_Eclipse is a big pack with a small dependency list (torch, numpy, Pillow, opencv-python, safetensors, pilgram, PyYAML, aiohttp) plus some heavier extras pinned for its audio/lyric nodes. If pip complains about imports, install into the Python that actually runs ComfyUI - for portable builds that means python_embeded\python.exe -m pip install -r .... Restart afterwards; nodes appear under 🌒 Eclipse/ Image.

Gotchas

Changing mode after you've connected an override does not rewrite the override text. The README says it plainly: update the string yourself, and don't assume a positive written for extract still makes sense in generate. Same for the file defaults on clothing, which are written to also fill in missing outfit pieces.

And treat extract as a lossy operation. You are asking a diffusion model to re-render a reference, and details it doesn't like will change - the author's own guidance is to preview the result before reusing it. If the source photo was good enough, reference mode costs you nothing and risks less.