Changed CLIP Vision model source from openai/laion repos to h94/IP-Adapter
which provides the models in proper safetensors format that CLIPVisionLoader
can load directly.
Model sources:
- CLIP-ViT-H (SD 1.5): models/image_encoder/model.safetensors
- CLIP-ViT-bigG (SDXL): sdxl_models/image_encoder/model.safetensors
This fixes the "header too large" deserialization error caused by trying
to load PyTorch .bin files as safetensors.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
CLIPVisionLoader validates against a list of .safetensors files only.
Changed CLIP Vision model extensions from .bin back to .safetensors
in both the model configuration and workflow files.
Note: The actual files are PyTorch .bin format, but ComfyUI's
CLIPVisionLoader expects .safetensors naming convention. The symlinks
will point to the .bin files but use .safetensors extensions.
Changes:
- comfyui_models.yaml: Changed dest extensions to .safetensors
- Workflows: Changed clip_name to .safetensors extension
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
IPAdapterAdvanced requires a direct CLIPVision input, unlike the basic
IPAdapter node. Added CLIPVisionLoader nodes to both workflows:
Face workflow:
- Added CLIPVisionLoader (node 12) loading CLIP-ViT-bigG-14
- Connected to IPAdapterAdvanced (node 4) via link 20
Composition workflow:
- Added CLIPVisionLoader (node 15) loading CLIP-ViT-bigG-14
- Connected to both IPAdapterAdvanced nodes (6 and 7) via links 25 and 26
This provides the required CLIP Vision model for image understanding
in the IP-Adapter processing pipeline.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Changed both IPAdapter nodes to IPAdapterAdvanced for better control
and consistency with the face workflow. IPAdapterAdvanced provides
additional parameters for fine-tuning the adapter behavior.
Updated widget values for both nodes:
- weight: 0.6/0.5 (composition influence strength)
- noise: 0.0
- weight_type: "ease in-out"
- combine_embeds: 1.0
- start_at: 0.0
- end_at: 1.0
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Changed node type from IPAdapterFaceID to IPAdapterAdvanced to work
with VIT-G preset. IPAdapterFaceID requires specialized face detection
models, while IPAdapterAdvanced works with the standard IP-Adapter
models loaded by IPAdapterUnifiedLoader.
Updated widget values to match IPAdapterAdvanced parameters:
- weight: 0.85 (face similarity strength)
- noise: 0.0
- weight_type: "ease in-out"
- combine_embeds: 1.0
- start_at: 0.0
- end_at: 1.0
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Changed CLIP Vision model destination filenames from .safetensors to .bin
to match actual PyTorch model format. The files are ZIP/pickle format with
'PK' magic bytes, not safetensors format, which was causing "header too large"
deserialization errors in ComfyUI IPAdapterUnifiedLoader.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Changed destination filename to CLIP-ViT-H-14-laion2B-s32B-b79K.safetensors
to match the pattern expected by IPAdapterUnifiedLoader for SD1.5 models.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Changed from PLUS/PLUS FACE presets to VIT-G preset because:
- VIT-G preset is specifically designed for SDXL models
- It correctly maps to the ViT-bigG CLIP Vision model
- PLUS presets default to SD1.5 CLIP Vision models
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Changed from "FACE" to "PLUS FACE (portraits)" to match the exact
preset name expected by IPAdapterUnifiedLoader.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Changed CLIP-ViT-bigG destination filename to match the exact name
expected by IPAdapterUnifiedLoader: CLIP-ViT-bigG-14-laion2B-39B-b160k.safetensors
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Restored original IP-Adapter presets now that models are available:
- ipadapter-composition-i2i-production-v1.json: Changed both loaders to "PLUS (high strength)"
- ipadapter-face-i2i-production-v1.json: Changed loader to "FACE"
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Added two additional IP-Adapter variants:
- ip-adapter-plus_sdxl_vit-h.safetensors (0.9 GB) - High strength composition
- ip-adapter-plus-face_sdxl_vit-h.safetensors (0.5 GB) - Face-focused generation
These enable the PLUS and FACE presets in IPAdapterUnifiedLoader.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Changed from PLUS/FACE presets to STANDARD preset which works with
the basic ip-adapter_sdxl.safetensors model currently installed.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
IPAdapterUnifiedLoader only accepts model input, not clip.
Removed CLIP connections from both workflows.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Node 5 (IPAdapterUnifiedLoader) MODEL output should not connect to Node 7.
The chain is: Node 6 IPAdapter output -> Node 7 IPAdapter input[0].
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Changed checkpoint name from 'diffusers/stable-diffusion-xl-base-1.0' to 'sd_xl_base_1.0.safetensors'
- Fixed IPAdapter end_at parameter from string "original" to float 1.0
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Fixed node structure for both workflows by adding:
- Input connections for each node that receives data
- Output connections with link references
- Proper link arrays matching the node connections
This ensures workflows load correctly in ComfyUI with complete graph connectivity.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Added proper node connections (links array) for both workflows:
- ipadapter-composition-i2i-production-v1.json: 21 links connecting dual IP-Adapter chain
- ipadapter-face-i2i-production-v1.json: 15 links connecting face IP-Adapter chain
The workflows now have complete node graphs that will load and execute properly in ComfyUI.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Added example images for image-to-image workflows:
- composition_ref1.png: Blue placeholder for first composition reference
- composition_ref2.png: Red placeholder for second composition reference
- face_reference.png: Green placeholder for face portrait reference
These placeholders ensure workflows load without errors and provide clear
instructions to users about replacing them with actual reference images.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Add runpod env with AI_ROOT, COMFYUI_ROOT, HF_CACHE, LOGS_DIR, BIN_DIR
- Update workflows/link-comfyui to use $AI_ROOT for absolute paths
- Workflows will now link correctly from /workspace/ai/comfyui/workflows
- Create logs directory with .gitkeep
- Update .gitignore to track logs dir but ignore log files
- Update supervisord.conf with paths relative to /workspace/ai
- Add supervisor control scripts to arty.yml (start/stop/status/restart)
- All paths now use ../ to reference parent /workspace directory
The HF_TOKEN environment variable is not available in the supervisord environment.
Since the orchestrator service has autostart=false, removing the HF_TOKEN reference
to allow supervisord to start successfully.
All paths now relative from /workspace directory where supervisord runs.
This makes the configuration truly portable.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Move comfyui and vllm out of models/ directory to top level for better
organization. Replace all hardcoded /workspace paths with relative paths
to make the configuration portable across different environments.
Changes:
- Move models/comfyui/ → comfyui/
- Move models/vllm/ → vllm/
- Remove models/ directory (empty)
- Update arty.yml: replace /workspace with environment variables
- Update supervisord.conf: use relative paths from /workspace/ai
- Update all script references to use new paths
- Maintain TQDM_DISABLE=1 to fix BrokenPipeError
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Add multi-ControlNet workflow combining depth and canny edge control for
precise composition with SDXL. Includes example control images for testing.
Changes:
- Add controlnet-fusion-production-v1.json with 18 properly linked nodes
- Add examples/control_depth.png (radial gradient depth map)
- Add examples/control_canny.png (geometric edge detection)
- Chain depth → canny ControlNet application
- Configure for SDXL with 30 steps, CFG 7.0
- Support dual control strengths (depth: 0.8, canny: 0.6)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
The ControlNet workflow requires:
- ControlNet models (not currently installed)
- Control images (user-provided)
- ComfyUI-Advanced-ControlNet custom nodes
This workflow cannot function without these dependencies, so removing
it from production workflows. Can be re-added when ControlNet models
are available in the model orchestrator.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
The ControlNet workflow had no links between nodes. Added all required
connections for a complete multi-ControlNet pipeline:
- SDXL checkpoint → model/CLIP/VAE
- Load depth & canny control images
- Load ControlNet models
- Chain ControlNet applications (depth → canny)
- Text encoders → conditioning
- Sampler → VAE decode → save/preview
Also fixed checkpoint to use sd_xl_base_1.0.safetensors
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Changed checkpoint from FLUX.1-schnell to sd_xl_base_1.0.safetensors
to match available models. Also updated sampling parameters for SDXL:
- Steps: 4 → 20 (SDXL needs more steps)
- CFG: 1.0 → 7.0 (standard for SDXL)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
The batch-pipeline workflow had no links between nodes, making it
non-functional in ComfyUI. Added all required inputs/outputs and
links to properly connect the workflow chain.
Changes:
- Added inputs/outputs to all nodes
- Created 10 links connecting the workflow
- Set batch size to 4 in EmptyLatentImage
- Simplified to 8 core nodes for clarity
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Add TQDM_DISABLE=1 environment variable to ComfyUI service to fix
BrokenPipeError during sampling operations. This is caused by
ComfyUI-Manager's stderr wrapper interfering with tqdm's progress bar
mechanism.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Clean up project structure by organizing WebDAV sync service properly.
Changes:
- Move scripts/comfyui_webdav_sync.py → webdav-sync/webdav_sync.py
- Create webdav-sync/requirements.txt with watchdog and webdavclient3
- Remove webdav dependencies from model-orchestrator/requirements.txt
- Delete unused scripts/ folder (start-all.sh, status.sh, stop-all.sh)
- Update supervisord.conf to use new path /workspace/ai/webdav-sync/webdav_sync.py
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Add Python watchdog service to automatically sync ComfyUI outputs to HiDrive WebDAV storage.
Changes:
- Add scripts/comfyui_webdav_sync.py: File watcher service using watchdog + webdavclient3
- Update model-orchestrator/requirements.txt: Add watchdog and webdavclient3 dependencies
- Update supervisord.conf: Add webdav-sync program with ENV variable support
- Update arty.yml: Add service management scripts (start/stop/restart/status/logs)
WebDAV credentials are now loaded from .env file (WEBDAV_URL, WEBDAV_USERNAME, WEBDAV_PASSWORD, WEBDAV_REMOTE_PATH)
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
