ComfyUI-Cluster

Ollama-driven routing nodes and curated workflows for ComfyUI.

Maintained by Geekatplay Studio — the console logs and node helpers now surface the Geekatplay branding so you can spot them quickly in shared setups.

This repo is cleaned to a focused setup:

• four production workflows
• checkpoint + LoRA flow
• flux split-component flow (diffusion model + VAE + text encoders)
• image-guided variants using Ollama vision analysis

model_registry.json is intentionally kept as the long-term reference registry.

Included Nodes

Exported custom nodes:

• OllamaPromptPlanner
• OllamaVisionStylePlanner
• DynamicCheckpointLoader
• DynamicLoraStack
• PreviewTextMerge
• LiveStatus

Workflows

All workflows are in workflows/:

• workflows/ollama_text_checkpoint_lora.json
• workflows/ollama_text_flux_split_lora.json
• workflows/ollama_image_checkpoint_lora.json
• workflows/ollama_image_flux_split_lora.json

Details are documented in docs/WORKFLOWS.md.

Install

1. Start Ollama and ensure your model is available (for example qwen2.5:7b and llava:7b).
2. Run:

install.bat

3. The installer will show all checkpoints/VAEs/text encoders/LoRAs from install_manifest.json. Pick what you want via the checkbox selector.

   New Features (v2):

   • Visual Progress Bar: Monitor BITS and WebClient downloads in real-time.
   • Civitai Support: Automatically resolves Civitai model page URLs to direct file links.
   • Smart Path Detection: Checks multiple ComfyUI folders (e.g., diffusion_models vs checkpoints) to avoid re-downloading existing models.
   • Robust & Resilient:
     • Auto-retries with fallbacks (BITS -> WebClient).
     • Detects and kills stalled downloads (60s timeout).
     • Automatically cleans up 0-byte corrupt files on failure.
   • Error Logging: Failed downloads are logged to install_errors.log for review.

4. If gated Hugging Face models are required, set HF_TOKEN environment variable before running.

Managing Models

If you manage your models in a custom directory (e.g., external drive), sync_registry.py supports interactive configuration and path repair.

python sync_registry.py --sync

• First Run: It will ask for your models root path if not found, and save it to cluster_config.json.
• Scanning: It checks checkpoints, diffusion_models, loras, etc.
• Repair: It fixes each registry item's path and folder_type when files moved or were registered incorrectly.
• Status Sync: It updates enabled based on whether the model file is currently found.
• Logging: A detailed scan report is written to model_scan.log.

Preview changes without writing:

python sync_registry.py --dry-run

Use manage_models.bat (wraps the python script) to quickly enable/disable nodes in the registry based on what you actually have installed.

Included manage_models.bat allows you to sync model_registry.json with your disk state:

• Sync: Repairs paths/types and disables models missing from disk.
• Reset: Re-enables all models in the registry.

Use this if you want the ComfyUI nodes to only offer models you actually have installed.

Source URL Repair (Download Locations)

If you need to validate/fix where models should be downloaded from (registry url fields), use sync_source_urls.py:

python sync_source_urls.py --checkpoints-only --check-live

This checks and repairs source URLs (Civitai/HuggingFace) instead of local installed paths.

Useful options:

• --dry-run: preview URL fixes without writing.
• --name-filter nightvision: target a single model or pattern.
• --checkpoints-only: only process checkpoint entries.
• --check-live: verify each final URL with HTTP checks.

The convenience launcher fix_missed_checkpoints.bat now runs source URL repair by default:

fix_missed_checkpoints.bat

Example targeted check:

fix_missed_checkpoints.bat --dry-run --name-filter nightvision --checkpoints-only --check-live

Installer behavior and model paths are documented in docs/INSTALL.md.

Notes

• Flux split workflows expect:
  • diffusion model in models/diffusion_models/
  • VAE in models/vae/
  • text encoders in models/text_encoders/
• LoRA loading uses models/loras/.
• DynamicCheckpointLoader fallback now respects per-model folder_type from model_registry.json.
• Console logging is prefixed with Geekatplay Studio | OllamaRouter to make routing messages easy to spot.