ComfyUI nodes for three/four-panel character cards, including a searchable visually curated 100-outfit wardrobe and 60-face library with live thumbnail previews, editable character controls and a persistent personal library.

Install

In ComfyUI Manager, use Install via Git URL: https://github.com/Coffmeteor/ComfyUI-MOSS-Card-Layout

Restart ComfyUI and refresh the browser. Existing users update this node pack to version 1.2.0 to get structured prompt fields and the aligned workflows. Registry/Manager search listing is pending; Git URL installation works independently of registry approval.

The prompt catalog ships with the node. Gallery images download automatically from https://ai-costume-gallery.pages.dev/ when their previews are requested, then remain in ComfyUI/user/moss-card-layout/images. Downloads are bounded, SHA-256 verified and atomically cached. No Hermes installation or private paths are required. First use needs internet; previously downloaded previews work offline. The images remain third-party assets under the source site's terms.

Use

Open Character Studio - Text.json from example_workflows. The top row contains aligned Face, Character/Card and Outfit sections. The actual final JSON, Result and a separate Advanced generation section sit below.

Click Choose face, search features or combine hair/face-shape tags, then click a thumbnail. This text route uses derived descriptions to generate a new face with similar traits; it does not preserve a reference person's identity.
Click Choose outfit, combine garment/color/material tags and select a thumbnail. Clothing fills the portrait and both clothed body views.
Set age, body shape, hairstyle, hair color and expression in Character Settings. From face selection keeps the selected description for that property. Explicit color/expression changes remove competing source details.
Choose dimensions, lighting and the privacy switch in Four-Panel Card Settings, then run. Change a component and run again to regenerate the card.

The component prompt contract is version 2. See field ownership and validation. New items with conflicting spatial fields fail before generation. The previous first-sentence clothing extractor is retired.

The requested panels are portrait, independent anatomy reference, clothed front and clothed back. Output acceptance remains unresolved: body panels can retain full heads or neck sections, anatomy panels can inherit shoes, and generated privacy shapes can differ from the requested rectangles. Review the native output; the component schema validates prompts, not these pixel-level requirements. The older Four View Visual Wardrobe workflow remains supported.

Add your own faces or outfits

Click Add my own, choose a PNG/JPEG/WebP (up to 16 MB and 32 megapixels), then enter a name, tags and short description. For faces, describe facial details, hairstyle, hair color and expression separately. For outfits, enter shared garment identity, front-only details, known back-only details and portrait clothing. Add asymmetric features with a wearer-side selector and front/back visibility; no JSON editing is required. Descriptions are entered manually; this node pack does not run a captioning model for uploaded images. The saved item is selected automatically. Use My library to find it later. Edit, Remove and Undo remove are available on your own items. IDs are internal and never needed for normal use.

User data lives in ComfyUI/user/moss-card-layout/library, outside the node package, and survives updates. Back up that directory with your workflows. Removal moves items into that library's .trash; the browser offers immediate undo. Workflows reference local library items: sharing a workflow JSON alone does not transfer personal images. Uploads stay on the ComfyUI server.

Models and reference editing

This package supplies selection, library and layout nodes. Text generation requires the existing MOSS HF integration, original SFT model and sixteen-step accelerator. Replace the example HF-directory placeholder MOSS-Image-SFT with your installed absolute HF directory and choose the installed accelerator in the generation section. It does not install model weights or replace their licenses.

The released workflow uses text conditioning. Experimental armV5 reference-card workflows are separate: local clothing edits can retain a reference face, while rebuilding a four-panel card can change facial features or body framing. Reference-card quality is being evaluated independently of this library package.

Face source

Face metadata comes from https://ai-portrait-gallery.pages.dev/. Images are fetched from the source on demand and cached under ComfyUI/user/moss-card-layout/portraits; they are not redistributed in this repository. Thumbnail hashes and full-image byte counts are checked. The 60 selected faces and 100 selected outfits were visually reviewed and manually rewritten into independent fields. Original reconstructed captions remain in source-provenance fields; these are not the source creator's original generation prompts. Source gallery terms apply.

Publishing

Publisher: coffmeteor. Registry publication uses the official Comfy-Org action. Create a publishing API key for this publisher in Comfy Registry, then store it as the GitHub repository Actions secret REGISTRY_ACCESS_TOKEN. Never commit it. Run Actions > Publish to Comfy Registry > Run workflow for the first publish. Later changes to pyproject.toml on main also trigger publication. Bump the semantic version for each new Registry release. Without the secret, the workflow reports that publication was skipped; a green setup run is not a published version.

See https://docs.comfy.org/registry/publishing for the official instructions.

Regional editing (1.3.0)

Requires native MOSS support in ComfyUI 0.43.0 or newer. The new nodes reuse an installed original MOSS HF snapshot and armV5; they do not download or duplicate model weights. MOSS HF Snapshot - Native resolves a snapshot name under ComfyUI/models/diffusers/ and exposes the native model, encoder and VAE.

Card Edit Region - JSON and Mask creates a real sampling mask from the same validated card region/bbox used by the JSON. Protect box edits around the box; turn it off to edit inside it. Card Region - Local Sampling crops the latent, mask and reference to exactly matching coordinates and dimensions. Card Region - Restore Canvas Latent restores protected and outside latent values before one complete-canvas VAE decode. No foreground segmentation or final-image face replacement is performed.

The face library now includes individually reviewed face rectangles for 60 portraits. Portrait Reference Placement transforms each rectangle through the same uniform center-crop/placement used for the whole reference photograph. Connect its suggested box to native Create Bounding Boxes to inspect/adjust it. The native interactive editor requires the Nodes 2.0 canvas setting. Personal uploads without reviewed metadata require a manually drawn protection box; the node refuses to guess one. Reference Character - Preserve Face and Expression inherits appearance and expression and exposes only body shape.

Keep the full-card generation JSON separate from the short local edit instruction. The Selected outfit mode derives only the selected view's garment description; it never copies the face geometry description into an editing instruction. Both positive and negative MOSS reference encoders must receive the exact local crop, without independently resizing it to a different megapixel target.

Protected-face editing retained the original facial features and expressions in three reviewed reference cases, including an identity outside the continuation training set, and in the integrated four-panel workflow. This preserves a face at its existing placement; it does not validate freely changing its pose or expression. Crown outpainting remains unreliable for cropped references. Native outputs can retain pale portrait-edge halos, neck remnants, and outfit/privacy rendering errors. The experimental continuation remains separate from the text workflow and is not bundled as a published model.