Character Reference Pack

If you have ever tried to keep a character's face, body and outfit straight across forty generated shots, you already know the failure mode: the face drifts, the jacket changes shade, and by image twelve you've invented a new person. Character Reference Pack [Eclipse] is the node that turns a pile of separate reference images into one written character definition plus a correctly numbered set of Qwen image slots. It generates nothing. It is pure prompt and wiring discipline, which is exactly what character consistency workflows are missing most of the time.

What it actually does

The node sits between your references and a Qwen image encoder. You feed it up to five images - face (required), plus optional body, clothing, rear and anchor - and it produces role instructions that tell the model which numbered image does what. The role prompts and the general negative come from prompts/character_references.json, which Eclipse extracts on first startup into its own repo folder. Edit that file and the next execution picks it up; no restart.

It has two lives, and mixing them up is the main way people confuse themselves:

• First pass (assembly). Leave anchor empty. Wire anchor_prompt to your encoder's positive, image_1…image_5 to the encoder's image inputs, and generate one clean full-body review character against the scene you set.
• Second pass (dataset). Connect that approved image into anchor, then send character_lock to Character Shot Planner instead of anchor_prompt. Now the review pose and the studio scene are gone, and what's left is the part you want repeated: identity, physique, wardrobe.

That distinction is the whole point of having two prompt outputs. anchor_prompt is the complete positive for the review shot; character_lock deliberately excludes the fixed pose and the scene so the planner can vary camera and pose without fighting your text.

Inputs worth setting

width and height (832×1216 by default, 256–2048 in multiples of 32) set the padded reference canvas. Images are padded, never stretched or cropped, so nothing about your subject gets mangled - but share the same dimensions between your assembly Pack and your dataset Pack or the numbering and framing drift.

Then it's prose, all multiline: identity (age, hair, distinctive features), body_description (relative proportions - chest-to-waist, limb lengths), outfit (fixed wardrobe, and explicit choices for missing pieces; clear the sample cardigan-and-jeans if you're using a clothing reference), scene (review image only), and extra_rules, which is appended to both prompts. Keep poses out of extra_rules if you want the planner to vary them.

face is marked optional in the widget but the node won't execute without it. body and clothing are genuinely optional - skip them and physique and wardrobe come from your written text instead.

Outputs

• character_lock → Character Shot Planner's character_lock (dataset phase).
• anchor_prompt → the assembly encoder's positive (first phase).
• negative → the same encoder's negative. The file's general assembly negative, separate from each Prepare node's per-role negative.
• image_1 … image_5 → the Qwen encoder's image inputs. These sockets have no fixed role. Active roles are packed consecutively: anchor if supplied, then face, body, clothing, rear.
• report → the actual image-number-to-role map, canvas size and omitted references. Send it to Show Any before you debug anything.

Install

ComfyUI Manager → search ComfyUI_Eclipse → install, or clone it yourself:

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

Portable installs should use python_embeded\python.exe -m pip install -r custom_nodes\ComfyUI_Eclipse\requirements.txt. Restart ComfyUI. The declared dependencies are modest - torch, numpy, Pillow, opencv-python, safetensors, pilgram, PyYAML, aiohttp - but requirements.txt also pins faster-whisper and stable-ts for Eclipse's lyric/transcription nodes. If you don't want that weight, install the small list by hand.

Where people get burned

Everything here is prompt text and a single image per input - batches are not supported, so a face sheet fed into face is one image, not five. And the reported role order is the only authority on what image_3 means; if you re-solve the graph by hand later, check report again rather than assuming. Finally, these are model instructions, not a guarantee: the README is blunt that role separation needs review, and the node doesn't measure BMI or infer concealed anatomy. On 2026 models the honest expectation is that an edit-model pass with a good reference does more for likeness than any amount of stacked adapter cleverness - pack the references properly here, then let Qwen-Edit do the heavy lifting.