12 Commits

Author SHA1 Message Date
Patrick Plate bf721c1379 feat(mcp): update bigmind/mcp-image-gen/webscraper servers; add image-gen batch scripts 2026-06-11 09:02:09 +02:00
Patrick Plate 1d8849cb41 fix(mcp-image-gen): confirmed working FLUX.2 Klein encoder filename
- CLIPLoader clip_name: qwen_3_4b_klein.safetensors (from Comfy-Org/vae-text-encorder-for-flux-klein-4b)
- VAE: flux2-vae.safetensors (321MB, same repo)
- Live test confirmed: 2.1MB photorealistic 1024x1024 PNG in 52.43s on RX 7900 XTX
- Test: assert clip_name == qwen_3_4b_klein.safetensors
- 37/37 tests pass
2026-04-10 20:29:18 +02:00
Patrick Plate 4a99a3625a fix(mcp-image-gen): rewrite flux2_klein_heretic workflow with CFGGuider + correct node types
- Replace FluxDisableGuidance+BasicGuider chain with CFGGuider (cfg=5)
- CLIPLoader: add device='default', keep type='flux2'
- UNETLoader: weight_dtype='default' (not fp8_e4m3fn — avoids dimension mismatch)
- VAEDecode/SaveImage: updated node IDs (11→VAEDecode, 12→SaveImage)
- Encoder: qwen_3_4b_bfl.safetensors (7.5GB BFL-merged shards)
- Tests: update heretic model assertions for new node structure (37/37 pass)
- Add RECAP doc with root cause analysis and session history
2026-04-10 20:21:12 +02:00
Patrick Plate ea0c5d39c4 fix(mcp-image-gen): fix Heretic/FLUX2 integration bugs
- Fix syntax error in server.py (dangling docstring lines)
- Correct model filename: flux-2-klein-4b.safetensors (without -fp8)
- Fix _WORKFLOW_REGISTRY key to match actual downloaded filename
- Update get_models() to always include registry models as fallback
- Fix test expectations to match corrected model names
- All 37 tests passing
2026-04-10 19:21:51 +02:00
Patrick Plate c662a5237b feat(mcp-image-gen): add ComfyUI auto-start health check + systemd service
Option A: Add lifespan context manager to server.py
- _ping_comfyui(): async health check against /system_stats
- check_and_start_comfyui(): ping on startup; if down, launches ComfyUI
  via subprocess.Popen from COMFYUI_DIR (.venv/bin/python main.py)
  with HSA_OVERRIDE_GFX_VERSION=11.0.0 injected for AMD ROCm
- Polls up to 30s for readiness after auto-start
- New env var: COMFYUI_DIR (default ~/ComfyUI)
- FastMCP lifespan= wired in; 34/34 tests still passing

Option B: Add comfyui.service systemd user service file
- Install: cp mcp/mcp-image-gen/comfyui.service ~/.config/systemd/user/
- Enable: systemctl --user enable --now comfyui
- Sets HSA_OVERRIDE_GFX_VERSION=11.0.0, WorkingDirectory=%h/ComfyUI
- Restart=on-failure, logs via journald

docs: Update mcp-image-gen-ComfyUI-Setup.md
- New Step 4: systemd service install + linger instructions
- Step 5: manual start (moved from old Step 4)
- Step 6/7 renumbered; COMFYUI_DIR env var documented
- Architecture diagram added; troubleshooting rows updated
2026-04-06 10:43:36 +02:00
Patrick Plate 79f1e6d65f feat(mcp-image-gen): add name and count params to generate_image
- Add name (str) param: filename prefix saved as {name}_{timestamp}_{seed}.png
- Add count (int, 1-10) param: generate N images in one call
- Extract _sanitize_name() helper: strips special chars, collapses underscores, caps at 64 chars
- Extract _build_filename() helper: pure function for testable filename construction
- Extract _generate_single() coroutine: clean loop body for batch generation
- Fixed seed batches increment seed per image (seed+i-1) for deterministic variation
- random seed (-1) batches give independent random seeds per image
- Partial batch failures continue (error TextContent in slot, remaining images proceed)
- Returns flat interleaved [Text1, Image1, Text2, Image2, ...] list
- 34/34 tests passing (was 19, added 15 new tests)
2026-04-06 07:45:37 +02:00
Patrick Plate d5510f590e Added new picture for bigmind page 2026-04-04 20:03:59 +02:00
Patrick Plate 64c0a62b49 feat(mcp-image-gen): add test suite (19 tests) and Lumen profile pictures 2026-04-04 14:09:11 +02:00
pplate 4165018ab2 fix(mcp-image-gen): fix HuggingFace authenticated download instructions
FLUX.1-schnell is a gated model — bare wget returns HTTP 401.

- Replace bare wget with huggingface-cli login + download (Option A)
- Add wget with Authorization header as Option B
- Add license acceptance prerequisite (huggingface.co gated repo)
- Add token creation link (huggingface.co/settings/tokens)
- Add fp8 quantized variant as alternative (~8.1GB, faster inference)
- Add download size note (~8GB, 10-30min)
2026-04-04 12:28:20 +02:00
pplate 2f01ff0639 fix(mcp-image-gen): correct ComfyUI install instructions in USAGE.md
ComfyUI is NOT on PyPI — `pip install comfyui` fails with
"No matching distribution found". Remove the wrong Option A.

Replace with:
- Warning note that pip install does not work
- Only correct method: git clone from GitHub + pip install -r requirements.txt

ROCm status confirmed: rocm-smi 3.1.0 / ROCm-SMI-LIB 7.7.0 installed.
2026-04-04 12:20:28 +02:00
pplate 8cbeb6571b docs(mcp-image-gen): add USAGE.md and expand tests to 19 2026-04-04 12:16:03 +02:00
Patrick Plate 8112ff2f12 feat(mcp-image-gen): scaffold ComfyUI-backed image generation MCP server
- FastMCP server with 4 tools: generate_image, list_available_models,
  get_generation_status, get_output_directory
- ComfyUI REST API client (httpx) polling lifecycle
- FLUX.1-schnell workflow JSON template
- Dual output: TextContent (path + seed) + ImageContent (base64 PNG)
- 14 passing pytest tests with respx HTTP mocking
- ROCm/AMD RX 7900 XTX optimized setup in README
- Ollama Linux migration path documented (future)
2026-04-04 11:49:31 +02:00