Character Sheet Pack

Building a character reference sheet by hand is a chore: blank mannequin, then the outfit on the mannequin, then a three-panel front/rear/portrait sheet, then three separate crops you can actually feed a dataset or a LoRA training run. Character Sheet Pack [Eclipse] is that chore as one node with a stage dropdown - four tasks that each expect different inputs and produce up to three numbered images plus a prompt.

The four stages, and why the inputs change per stage

stage is the first widget and it's the one that matters:

• mannequin - generate a blank body to dress later. Two modes here: generation_mode=Preset skips every image input and builds the figure from the body dropdowns (gender, build, muscularity, breast_size, chest_breadth, hip_width, buttock_size) plus a blank size-setting canvas; Reference ignores the presets and uses front_body (and optionally rear_body) as physique photographs. The README is careful about this: those body presets are qualitative prose instructions, not measurements, and unspecified adds nothing rather than guessing.
• wardrobe - garment-only product views, from an outfit source photo. Wearers and mannequin supports are meant to be excluded; prints and openings on the source garment are preserved, as is whatever is visible through sheer fabric.
• sheet - the actual three-panel character sheet. Needs the approved mannequin guide in image_1 and a face portrait, plus an optional wardrobe board. If you turn off use_wardrobe, you must supply a complete written outfit in details.
• dataset - individual approved panels: front plus face, optionally rear.

The image numbering follows the stage and compacts. On the mannequin stage with only a rear physique photo supplied, that photo is image_2, not image_3. Read report before you trust a slot: it lists the actual label-to-role mapping, the omitted references and both the target canvas and each reference's real dimensions.

Inputs you'll touch

width/height (1536×864 by default) set the canvas, and the Qwen encoder must run at resolution 0 for that canvas size to be honoured. In the mannequin and sheet stages only the first image gets fitted to the canvas; supporting face and wardrobe references keep their aspect ratio, are never upscaled, and are capped at 1024px on the long side. Other stages pad without stretching.

details is the free-text field that changes job per stage: physique for mannequin, missing outfit pieces for wardrobe, identity and outfit continuity for sheet and dataset. The four toggles - use_front_body, use_rear_body, use_wardrobe, use_rear - are per-stage and skip their input branch when off, so an unrelated image left connected to the wrong socket gets ignored rather than silently polluting the prompt.

positive_override and negative_override replace the stage text from prompts/character_sheets.json, but the node still resolves roles and appends details. Keep uppercase role names like FACE, WARDROBE, FRONT BODY in any custom prompt - the Pack rewrites those to the real <imageN> labels.

Outputs

positive and negative go to the same encoder; on the dataset stage, positive is the thing you send to Character Shot Planner's character_lock. image_1/image_2/image_3 go to that same encoder's image inputs - unused slots return nothing. report is your map, and the reason to wire up Show Any.

Install

ComfyUI Manager → ComfyUI_Eclipse, or:

cd ComfyUI/custom_nodes
git clone https://github.com/r-vage/ComfyUI_Eclipse
pip install -r ComfyUI_Eclipse/requirements.txt

Restart, then find it under 🌒 Eclipse/ Image. You'll also need a Qwen image encoder node - Eclipse's loaders and encoders moved to the sibling ComfyUI_SmartModelLoader pack, so install that too if a red missing-node box shows up on load.

Realistic expectations

This is a multi-pass pipeline at 1536×864 with up to three references. On 8–12GB you want the quantized/lightning Qwen-Edit path (nunchaku 2509 Lightning is the usual advice) or you'll be watching a very slow sampler. And don't expect free wins: sheets generated at 3 panels still need inspection - the sheet prompt asks for specific framing, but it can't enforce it, so if a divider lands in the wrong place or a figure goes missing, fix it before feeding the crops into a dataset. Generation instructions are not a geometric constraint, and the node says so itself.