Ethanfel
1ba04c6d88
Proxy JS widget files from the remote instance via /extensions/ComfyUI-Lora-Manager/ prefix, and handle send_sync routes locally instead of falling through to the original package. This eliminates the requirement to install ComfyUI-Lora-Manager alongside. - Add /extensions/ComfyUI-Lora-Manager/ to proxy prefixes - Replace _SEND_SYNC_SKIP_ROUTES with local handler functions that fetch data from remote and broadcast events via PromptServer.send_sync() - Add lm_remote_bootstrap.js to load Vue widget bundle from remote - Update docstrings and README to reflect standalone operation Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
145 lines
6.3 KiB
Markdown
145 lines
6.3 KiB
Markdown
# ComfyUI-LM-Remote
|
|
|
|
Remote-aware [LoRA Manager](https://github.com/willmiao/ComfyUI-Lora-Manager) nodes for ComfyUI. Fetches metadata (trigger words, hashes, model info) from a remote LoRA Manager instance via HTTP, while loading LoRA files from local NFS/SMB-mounted paths.
|
|
|
|
## Why?
|
|
|
|
When ComfyUI runs on a GPU workstation and LoRA Manager runs in Docker on a NAS (e.g., Unraid), the original LoRA Manager nodes can't access the remote metadata database. This package bridges that gap:
|
|
|
|
- **Proxy middleware** transparently forwards the LoRA Manager web UI and API to the remote instance
|
|
- **Remote nodes** fetch metadata via HTTP instead of local SQLite lookups
|
|
- **Local file loading** is unchanged -- LoRA files are loaded from shared storage (NFS/SMB)
|
|
|
|
```
|
|
ComfyUI Workstation NAS (Docker)
|
|
+--------------------------+ +------------------------+
|
|
| ComfyUI | | LoRA Manager |
|
|
| +- ComfyUI-LM-Remote | HTTP API | +- SQLite metadata DB |
|
|
| | (this package) |<------------>| +- CivitAI sync |
|
|
| +- /mnt/loras/ (NFS) | | +- Port 8188 |
|
|
+--------------------------+ +------------------------+
|
|
| |
|
|
+------- Shared NFS/SMB storage ------------+
|
|
```
|
|
|
|
## Prerequisites
|
|
|
|
- A running LoRA Manager instance accessible over the network (e.g., in Docker)
|
|
- Shared storage so both machines see the same LoRA files at compatible paths
|
|
|
|
> **Note:** The original [ComfyUI-Lora-Manager](https://github.com/willmiao/ComfyUI-Lora-Manager) package is **not required**. Widget JS files and Vue widget types are served from the remote instance via the proxy. You may still install it alongside if you want the original (non-remote) nodes available too.
|
|
|
|
## Installation
|
|
|
|
Clone into your ComfyUI `custom_nodes/` directory:
|
|
|
|
```bash
|
|
cd /path/to/ComfyUI/custom_nodes/
|
|
git clone https://github.com/ethanfel/ComfyUI-LM-Remote.git
|
|
```
|
|
|
|
## Configuration
|
|
|
|
Edit `config.json` in the package directory:
|
|
|
|
```json
|
|
{
|
|
"remote_url": "http://192.168.1.3:8188",
|
|
"timeout": 30,
|
|
"path_mappings": {}
|
|
}
|
|
```
|
|
|
|
| Field | Type | Default | Description |
|
|
|-------|------|---------|-------------|
|
|
| `remote_url` | string | `""` | URL of the remote LoRA Manager instance |
|
|
| `timeout` | int | `30` | HTTP request timeout in seconds |
|
|
| `path_mappings` | object | `{}` | Remote-to-local path prefix mapping (see below) |
|
|
|
|
### Environment Variable Overrides
|
|
|
|
Environment variables take priority over `config.json`:
|
|
|
|
| Variable | Overrides |
|
|
|----------|-----------|
|
|
| `LM_REMOTE_URL` | `remote_url` |
|
|
| `LM_REMOTE_TIMEOUT` | `timeout` |
|
|
|
|
### Path Mappings
|
|
|
|
If the remote instance and local ComfyUI see LoRA files at different absolute paths, use `path_mappings` to translate:
|
|
|
|
```json
|
|
{
|
|
"path_mappings": {
|
|
"/data/models/loras": "/mnt/nas/models/loras"
|
|
}
|
|
}
|
|
```
|
|
|
|
This maps the remote path prefix `/data/models/loras` to the local `/mnt/nas/models/loras`. Usually not needed if both machines use the same NFS mount point.
|
|
|
|
## Nodes
|
|
|
|
All nodes appear under the **Lora Manager** category in the ComfyUI node menu, with "(Remote, LoraManager)" in the name.
|
|
|
|
| Node | Description |
|
|
|------|-------------|
|
|
| **Lora Loader (Remote)** | Load LoRAs with trigger words from remote metadata. Supports Nunchaku Flux models. |
|
|
| **LoRA Text Loader (Remote)** | Load LoRAs from `<lora:name:strength>` text syntax. |
|
|
| **Lora Stacker (Remote)** | Stack multiple LoRAs into a LORA_STACK for downstream loaders. |
|
|
| **Lora Randomizer (Remote)** | Randomly sample LoRAs from the remote pool with configurable count and strength ranges. |
|
|
| **Lora Cycler (Remote)** | Sequentially cycle through LoRAs from the remote pool. |
|
|
| **Lora Pool (Remote)** | Configure pool filters (base model, tags, folders, favorites) for Randomizer/Cycler. |
|
|
| **Save Image (Remote)** | Save images with embedded generation metadata, using remote hash lookups for LoRA and checkpoint hashes. |
|
|
| **WanVideo Lora Select (Remote)** | Select LoRAs for WanVideo with block-level control. |
|
|
| **WanVideo Lora Select From Text (Remote)** | Select WanVideo LoRAs from text syntax. |
|
|
|
|
## How It Works
|
|
|
|
### Reverse Proxy
|
|
|
|
An aiohttp middleware is registered at startup that intercepts requests to LoRA Manager endpoints and forwards them to the remote instance:
|
|
|
|
**Proxied routes:**
|
|
- `/api/lm/*` -- all REST API endpoints (except send_sync routes below)
|
|
- `/extensions/ComfyUI-Lora-Manager/*` -- widget JS files and Vue widget bundle
|
|
- `/loras_static/*`, `/locales/*`, `/example_images_static/*` -- static assets
|
|
- `/loras`, `/checkpoints`, `/embeddings`, `/loras/recipes`, `/statistics` -- web UI pages
|
|
- `/ws/fetch-progress`, `/ws/download-progress`, `/ws/init-progress` -- WebSocket connections
|
|
|
|
**Handled locally** (events broadcast to local browser via `send_sync`):
|
|
- `/api/lm/loras/get_trigger_words` -- fetches trigger words from remote, broadcasts `trigger_word_update`
|
|
- `/api/lm/update-lora-code` -- broadcasts `lora_code_update`
|
|
- `/api/lm/update-node-widget` -- broadcasts `lm_widget_update`
|
|
- `/api/lm/register-nodes` -- no-op in remote mode
|
|
|
|
### Remote Metadata
|
|
|
|
Nodes use `RemoteLoraClient` to fetch metadata from the remote LoRA Manager API. A 60-second in-memory cache avoids redundant API calls during workflow execution. The client queries:
|
|
|
|
- `GET /api/lm/loras/list` -- full LoRA list (cached)
|
|
- `GET /api/lm/loras/get-trigger-words` -- trigger words fallback
|
|
- `POST /api/lm/loras/random-sample` -- random sampling
|
|
- `POST /api/lm/loras/cycler-list` -- cycler ordering
|
|
- `GET /api/lm/checkpoints/list` -- checkpoint hashes (cached)
|
|
|
|
### Local File Loading
|
|
|
|
After fetching the relative path from the remote metadata, LoRA files are loaded locally via `folder_paths.get_full_path("loras", relative_path)`. No files are transferred over the network.
|
|
|
|
## Verification
|
|
|
|
After installation and configuration:
|
|
|
|
1. Restart ComfyUI
|
|
2. Check logs for: `[LM-Remote] Proxy routes registered -> http://192.168.1.3:8188`
|
|
3. Open ComfyUI -- the LoRA Manager web UI should load (proxied from remote)
|
|
4. Add a "Lora Loader (Remote, LoraManager)" node to a workflow
|
|
5. Select a LoRA -- trigger words should populate from remote metadata
|
|
6. Run the workflow -- the LoRA loads from local shared storage
|
|
|
|
## License
|
|
|
|
Same license as [ComfyUI-Lora-Manager](https://github.com/willmiao/ComfyUI-Lora-Manager).
|