Ultimate Lora Loader

A dynamic LoRA-stack loader for ComfyUI, in the spirit of rgthree's Power Lora Loader, with one core difference: clicking Add Lora opens a real folder browser instead of one giant flat list. Click into subfolders, breadcrumb back out — nested files show up where they actually live on disk instead of as subfolder/name.safetensors strings buried in an alphabetical dropdown.

The node in its default state, right after adding it to the canvas:

Features

• Folder-aware LoRA browser — the Add Lora popup mirrors your actual models/loras folder structure, with clickable breadcrumbs to navigate back up
• Optional CLIP — leave clip unconnected to apply LoRAs to the model only (same underlying behavior as ComfyUI's own LoraLoaderModelOnly); connect it and a second, independent CLIP-strength column appears automatically
• Per-row split strengths — model and clip strength are tracked and adjustable independently, not mirrored
• Drag-to-reorder — grab a row by its handle to reorder the stack, since LoRA application order can affect results
• Priority field — type a target position directly instead of dragging, if you prefer
• Native quick-edit dialogs — click any strength or priority number to get ComfyUI's own auto-selected value-entry dialog (matches the interaction rgthree's nodes use)
• Toggle-all / delete-all — bulk controls in the header row, column-aligned with the per-row toggle and trash icons
• Auto-growing node — the node resizes itself to fit your lora list; shrink it manually below that and the list scrolls internally instead of spilling out
• Missing-file detection — if a row points at a lora that's no longer on disk (moved, renamed, or deleted since the workflow was saved), the row is flagged in red with a ⚠ marker, and its name becomes click-to-replace so you can swap in a working file without rebuilding the row

A populated stack — drag handle, priority number, toggle, name, strength stepper, and trash icon, all column-aligned with the header row's toggle-all and delete-all controls:

Disabling a row (toggle off) dims it, while an added third row still keeps the enabled ones visually distinct:

Connect clip and every row gains an independent second stepper — the header switches from a single "Strength" column to "Model" and "Clip":

If a row's lora file goes missing — moved, renamed, or deleted since the workflow was saved — that row is flagged in red with a ⚠ marker instead of silently failing at generation time. Click the highlighted name to reopen the folder browser and pick a replacement in place:

Install

Via git clone (recommended):

cd ComfyUI/custom_nodes
git clone https://github.com/WaitWut/comfyui-ultimate-lora-loader.git

Manually: download this repo as a ZIP and extract it into ComfyUI/custom_nodes/ so you end up with:

ComfyUI/custom_nodes/comfyui-ultimate-lora-loader/
├── __init__.py
├── nodes.py
├── pyproject.toml
├── LICENSE
├── README.md
└── js/
    └── ultimate_lora_loader.js

Then restart ComfyUI completely (not just a browser refresh — new custom nodes only load on a fresh process start).

ComfyUI Desktop users: the actual custom_nodes path may not be where you'd expect. Check Settings → System Paths in the app to find it — custom nodes have occasionally landed in the Electron app's own resources folder rather than the data directory you originally chose during setup.

No extra Python dependencies — this only uses folder_paths, comfy.sd, comfy.utils, and aiohttp, all of which ship with ComfyUI already.

Usage

1. Add the node: right-click canvas → loaders → Ultimate Lora Loader, or double-click the canvas and search "Ultimate Lora Loader".
2. Wire model in. Wiring clip is optional — see below.
3. Click Add Lora. Browse into folders, click a .safetensors/.pt file to add it as a row.
4. Click a strength number (or the priority number) to open a native quick-edit dialog, pre-selected for typing a new value — or use the ◀ ▶ arrows to nudge by 0.05 at a time.
5. Drag a row by its handle (⋮⋮ icon on the left) to reorder the stack, or type a new number directly into the priority field to jump a row to that position.
6. Outputs feed straight into your sampler/conditioning chain like normal.

CLIP is optional

If you don't connect clip, the node applies each LoRA to the model only (mirroring ComfyUI's own LoraLoaderModelOnly node internally), and the per-row Clip-strength column simply doesn't appear. Connect clip at any point and the column shows up immediately; disconnect it and it goes away — no need to re-add your loras either way.

Does lora order matter?

Yes. Each LoRA in the stack is applied as a sequential patch on top of whatever the previous one already changed, so which one goes first can affect the result — usually a smaller effect than strength values, but real, especially with multiple LoRAs touching overlapping weights at higher combined strengths. That's what the drag handle and priority field are for.

How the folder tree works

• Backend: GET /ultimate_lora_loader/tree walks folder_paths.get_filename_list("loras") (so it automatically respects any extra search paths you've set in extra_model_paths.yaml) and re-nests the flat subfolder/file.safetensors strings into a real nested dict by splitting on /.
• Frontend caches that tree for 15 seconds per popup-open cycle so repeated clicks don't hammer the endpoint, but refreshes automatically after that — new LoRAs dropped into a folder on disk show up within ~15s without needing a full ComfyUI restart.

Known limitations

• The stored lora identity in each row is the relative path string (e.g. characters/lora1.safetensors), same as vanilla ComfyUI lora loaders use internally — so workflows/JSON stay portable across machines as long as the same relative folder structure exists.
• No search/filter box in the Add Lora popup yet — for very large flat folders you'd still be scrolling. Fine for a folder-per-character/style setup, less fine if you dump 500 loras in one subfolder. (Planned — see Roadmap below.)

Roadmap

Planned features, roughly in the order they're likely to land:

Self-contained (no external services needed):

• Search/filter box in the Add Lora popup — type to filter instead of scrolling through folders
• Duplicate detection — block re-adding a lora that's already in the stack, with a clear message, since there's no real use case for having the same file twice

Trigger words, suggested strength, and CivitAI metadata (one combined feature):

• A clickable info icon after the lora name, showing trigger words and CivitAI metadata for that lora
• Suggested strength range pulled from CivitAI where available — this isn't stored in the .safetensors file itself, so it depends on the CivitAI lookup rather than being derivable locally

Save/load presets:

• Save a particular lora + strength stack as a named preset, and re-apply it later without rebuilding it lora-by-lora

Have a feature idea not listed here, or want to help build one of these? Open an issue — see below.

Contributing

Bug reports, feature requests, and PRs are all welcome via GitHub Issues. If you're planning a larger change, opening an issue first to discuss it is appreciated but not required.

Credits

Inspired by rgthree-comfy's Power Lora Loader — this project doesn't reuse rgthree's code, but the row-based UI pattern and native quick-edit dialog behavior were built to match it closely on purpose, since it's a well-established interaction users already know.

License

MIT