Model Gallery

comfyui-model-gallery

Touch-first card-grid picker with preview thumbnails for the folder-backed model combos (LoRA, checkpoint, VAE, ControlNet, UNet, CLIP, upscale).

Part of a family of mobile-first ComfyUI usability packs built on the shared @laurigates/comfy-modal-kit — gallery-loader, prompt-editor, sampler-info, touch-numeric, touch-connect, touch-resize, touch-tooltips: touch-friendly HTML modals that replace clunky native LiteGraph controls, detected by widget name, additive and non-clobbering.

Via the kit's cross-pack field-provider registry (v0.4.0) this pack also registers the gallery as an inline control, so an editor built on the kit (e.g. prompt-editor) can mount the model picker directly in a field row — not only on canvas tap. See ADR-0002.

The card-grid picker over a ckpt_name combo: subfolder filter chips, fuzzy name filter, and the current value highlighted. (Screenshot uses placeholder model names.)

Install

cd <ComfyUI>/custom_nodes
git clone https://github.com/laurigates/comfyui-model-gallery

Restart ComfyUI; hard-refresh the browser tab (Ctrl+Shift+R / Cmd+Shift+R).

What it does

Detects folder-backed model combos by widget name (ckpt_name, lora_name, vae_name, control_net_name, unet_name, clip_name*, upscale_model, …) and replaces the native LiteGraph dropdown with a touch-first card grid: subfolder filter chips, fuzzy name filter, mtime, and the current value highlighted. The exact combo string is written back verbatim so serialized workflows never churn.

Model info

Each card is annotated with what the model is, in two layers:

1. Filename corpus (instant, offline, every file type) — base-architecture family + notable-model badges and a one-line summary, matched by pattern against the filename (web/data/models.json). Works for .ckpt/.pth/GGUF and files with no embedded metadata. The same info also feeds the fuzzy filter (search sdxl, anime, upscale, …) and the widget's hover/long-press tooltip for the current value.
2. Embedded metadata (authoritative, on demand) — tap a card's ⓘ to read the .safetensors header (base model, precision/params, rank/alpha, trained resolution, and the most-frequent training tags). Served by the /model_gallery/meta backend endpoint, which parses only the file header (no tensors) using bundled libs and resolves paths solely through folder_paths — it never reads an arbitrary path.

The corpus is heuristic — a hint, not a guarantee; embedded metadata wins when present. Both are additive: a file with no match just shows its bare name.

Where the grid shows up

Three surfaces, in increasing indirection — all optional, all falling back cleanly to the native control when a sibling pack isn't installed:

1. On the canvas. Tapping a matched *_name combo opens the grid as a modal.
2. Inline in a node editor. The pack registers a FieldProvider with @laurigates/comfy-modal-kit, so comfyui-prompt-editor's all-fields modal mounts the grid in place of a bare <select> for those same combos.
3. Anywhere a host knows a folder category but has no combo widget. The pack also registers a ModelPicker, keyed on the folder_paths category (loras, checkpoints, …) rather than on a widget name.

The third exists for rgthree's Power Lora Loader: its LoRA rows are self-drawing custom widgets whose value is an object with no values array, so they match neither of this pack's widget-name predicates. A host editing such a row can now ask for "a picker for loras" and get the full card grid, plus an optional compact metadata strip (base architecture, rank/alpha, copy-on-tap trigger words) for the file already chosen. All metadata knowledge stays inside this pack — the host mounts an opaque element.

The set of pickable categories is derived from the widget-name map itself, so adding a widget mapping automatically makes its category pickable and the two cannot drift.

LoRA details

LoRAs carry more in their header than any other model type, and the detail fold surfaces it:

• Trigger words — the tokens the LoRA was captioned with (ss_trained_words / modelspec.trigger_phrase), as chips you tap to copy straight into a prompt, plus Copy all. They also appear in the widget's hover/long-press tooltip for the selected LoRA, so you can read them without opening the picker. Frequency-ranked dataset tags (ss_tag_frequency) are shown separately below them — a statistic, not a declared trigger.
• Topology — rank, alpha and the effective weight scale (α / r), the adapter implementation (networks.lora, lycoris.kohya, …), CLIP skip, and the base checkpoint it was trained against.
• Training — optimizer, learning rates (UNet / text encoder), steps, epochs, dataset image counts and aspect-ratio bucket count.
• Civitai — a link to the model page when a download helper injected civitai_model_id / civitai_version_id. The link is yours to click; the pack itself makes no outbound request.

All of this is optional in the file format: LoRAs trained with kohya/sd-scripts or Musubi Tuner carry most of it, while raw diffusers or minimalist scripts often carry none. Missing fields are simply omitted — never guessed.

Compatibility

• ComfyUI: modern Vue frontend (comfyui-frontend-package >= 1.40) for the widget.onPointerDown interception hook.
• Frontend changes (JS/CSS) take effect on browser hard-refresh — no restart.

License

MIT — see LICENSE.