comfyui-mcp

Drive the LTX Director (Timeline) node — its Add Image/Text/Audio buttons are DOM-only and cannot be clicked by an agent; edit the hidden timeline_data JSON widget instead. Load when a workflow contains LTXDirector / LTXDirectorGuide / PromptRelayEncodeTimeline, or when asked to add, move, retime, or remove timeline segments (image / text / audio / motion).

Train a character/identity LoRA locally on FLUX.1-dev via the comfyui-mcp train_* tools (GPU Docker + ostris ai-toolkit). Use when the user wants to train a LoRA of a person/character from their photos on the local GPU — covers dataset prep, launch, monitoring, and using the result in ComfyUI. For WAN/Z-Image training via the ai-toolkit UI see ai-toolkit-trainer.

Core ComfyUI knowledge — workflow format, node types, pipeline patterns, and MCP tool usage

Discover Civitai models with the BUILT-IN search_civitai_models tool and install/generate them locally — find a checkpoint/LoRA/embedding on Civitai, download it into ComfyUI, and use its trigger words. Optionally pair the official Civitai MCP for community features (images browsing, posting, collections).

Run the ComfyUI agent locally for FREE — no subscription, no API key, fully offline — using our gemma4 models fine-tuned on the comfyui-mcp tool suite via Ollama. Use when the user asks about running locally, running for free, offline use, avoiding API costs, Ollama setup, or which local model to pick.

Debug a WRONG or imperfect render (not a hard error) by inspecting inputs and intermediate steps with run-to-node — render one branch up to an output, preview-tap latents/masks/preprocessor maps, localize the first bad stage, then fix. Use when a final image/video comes out wrong — artifacts, wrong subject/pose/composition/color, blur, a ControlNet/IPAdapter/mask/LoRA not taking, a two-stage refiner or upscale degrading the result — rather than the run failing with an error (for errors/OOM/missing nodes use the troubleshooting skill).

Full production pipeline — story to scenes, Z-Image start frames, Qwen Edit end frames, WAN FLF video clips, ffmpeg concatenation

Lay out and organize a ComfyUI workflow cleanly on the live panel canvas — dependency-layered node placement with no overlaps, subgraphs, colored group boxes, and subgraph rail alignment. Use when asked to tidy / clean up / organize / arrange a workflow, add groups or subgraphs, fix overlapping nodes, or build a workflow that should look good from the start.

Build Ideogram 4 (Ideogram Ultra) txt2img and img2img workflows — local open-weights model, dual conditional/unconditional models with DualModelGuider, Qwen3-VL text encoder, and structured JSON ("compositional deconstruction") prompts for strong text rendering and layout control

Pick the right ComfyUI startup flags for VRAM, attention, caching, and speed — the full decision matrix for OOM (--novram / --cache-none / --disable-smart-memory), shared-VRAM creep on Windows (--reserve-vram N), model-switching with big text encoders (--cache-none), high-VRAM throughput (--gpu-only / --highvram), and attention-backend selection (--use-sage-attention for speed, --use-pytorch-cross-attention as the highest-quality / Z-Image-safe fallback). Also the acceleration-stack + Blackwell/RTX 5000 (sm_120) notes. Use when a graph OOMs (especially long video like LTX 2 / WAN), when the GPU spills into shared VRAM and slows to a crawl, when switching between models eats all RAM, when Z-Image produces black/garbled output under Sage, or when deciding which attention backend to launch with. Flag names verified against upstream comfy/cli_args.py — see Sources.

Install Triton + SageAttention to accelerate ComfyUI (the sageattn attention_mode and inductor torch.compile used by WanVideoWrapper / many video graphs) — Windows-first (triton-windows + woct0rdho prebuilt SageAttention wheels matched to torch/CUDA/python into the RIGHT python), plus Linux (official triton + build) and Mac (N/A → sdpa/MPS). Critically also covers the SAFE sdpa / no-compile fallback so an example that assumes sageattn + torch.compile still runs when these aren't installed (video-extend TRAP 5). Use when a loader crashes with "No module named 'sageattention'" / "triton: unavailable", when asked to speed up Wan/video workflows, or when deciding whether to install acceleration vs. fall back.

Common ComfyUI errors and fixes — OOM, missing nodes, dtype mismatches, black images, and debugging strategies

Build Z-Image txt2img workflows — RedCraft checkpoint, Z-Image Turbo/Base LoRAs, ControlNet, and sampler presets

Build WAN MultiTalk audio-driven talking-avatar / lip-sync video workflows — MeiGen-AI MultiTalk on WAN 2.1 14B I2V via kijai WanVideoWrapper (portrait + audio → lip-synced video)

Diagnose and fix video/image color OBJECTIVELY with the analyze_color tool (scopes/stats — black/white points, contrast, saturation, clipping, cast) instead of eyeballing a contact sheet. Covers the "washed out" signature, why reference color-match (mkl/ColorMatch/ColorMatchAdobe) CAN'T add contrast a flat source lacks, the levels/contrast-stretch fix (core AdjustContrast / CurveEditor), the measure→fix→re-measure loop, the side-by-side sandbox pattern, and where to place the fix in a render graph (after decode, before save). Use when a render looks washed out / flat / dull / over-saturated / color-cast, or when deciding between a color-match and a contrast/levels fix.

Build Krea 2 Turbo txt2img workflows — native krea2 CLIPLoader, Qwen3-VL encoder, Qwen image VAE, 8-step turbo settings, and Ideogram-style JSON prompting

Build Lightricks LTX-2 / LTX-2.3 video workflows — text-to-video, image-to-video, GGUF and bundled checkpoints, distilled model, camera control LoRAs, synchronized audio, two-stage upscaling, and swapping alternate/GGUF base models

Extend / continue a video temporally with Pusa 2.2 in ComfyUI — temporal flowmatching (the flowmatch_pusa scheduler + WanVideoAddPusaNoise) on the WanVideoWrapper stack with WAN 2.2 T2V A14B (HIGH/LOW) models and the Pusa V1 LoRAs, conditioning on the loaded clip via WanVideoEncode so the existing motion carries into the continuation. Covers the kijai wanvideo_2_2_14B_Pusa_extension graph, model/LoRA slots + downloads, noise/length/scheduler settings, chaining multiple extensions, VRAM tiers, gotchas, and the extend→upscale handoff.

Build WAN 2.2 First-Last-Frame video workflows — native dual hi-lo (required), and WanVideoWrapper VACE approaches

Upscale and restore video in ComfyUI — both the quick local path (per-frame ESRGAN like 4x_foolhardy_Remacri via ImageUpscaleWithModel + 4x→2x supersample, with its temporal-flicker tradeoff) and temporal-aware super-resolution (SeedVR2, the newer FlashVSR) with the downscale-first restore pipeline; RIFE/FILM frame interpolation via the BUILT-IN ComfyUI 0.26 FrameInterpolate (rife_v4.26 in models/frame_interpolation/) or the ComfyUI-Frame-Interpolation pack; 2x/4x scaling, VRAM tiers, VHS encode. Captures the classic downscale→SeedVR2→RIFE recipe and the current 2026 recommendation.

Build WAN 2.2 Text-to-Video workflows — dual hi-lo models, lightning LoRAs, VACE modules, and KSamplerAdvanced two-pass

Self-heal + report bugs to make the ecosystem better. When you hit a reproducible bug in comfyui-mcp / the sidebar panel, OR in a third-party custom node / ComfyUI core, use this: diagnose it, attempt a local fix so the user can keep working, and report it the right way (our intake Worker for our repos; the node's own GitHub for third-party). Acts autonomously and just informs the user — it does not ask permission for each step. Triggers on tool/panel/orchestrator errors you can't work around, or when the user says "report this" / "fix this bug".

Build Baidu ERNIE-Image / ERNIE-Image-Turbo workflows — primarily TEXT-TO-IMAGE. Pick ERNIE when you need precise multilingual text rendering, posters/signage, manga/anime multi-panel layouts, or strong instruction following for complex multi-object scenes. Also supports denoise-based image-to-image refine (NOT instruction-grounded editing — use Qwen-Image-Edit or Flux Kontext for "change X in this photo" edits).

Build Flux txt2img workflows — Flux.1 Dev (SRPO), Flux 2 Klein 9B, Turbo LoRAs, FluxGuidance, and DualCLIPLoader patterns

Train custom LoRAs with ostris AI-Toolkit — covers WAN 2.2/2.1 (people, styles, video motion) and Z-Image (Turbo & Base, low-VRAM image LoRAs). Use when the user wants to train a WAN or Z-Image LoRA; covers local + RunPod setup, dataset prep, key params, and using the result in a ComfyUI workflow.

Use when installing a model family from an installer pack, or when building/deriving a new pack from an upstream installer or a workflow JSON. Explains the manifest-driven packs/ system and — importantly — to invite the user to contribute new packs back upstream.

Anime/illustration text-to-image (ANIMA 1.0, ~2B Cosmos DiT) — use for anime, manga, illustrated characters; accepts Danbooru tags + natural language; runs/trains on <6GB VRAM; includes anime inpainting via Anima-LLLite ControlNet

Train a custom anime LoRA on the ANIMA base model — Citron's local Gradio trainer (kohya sd-scripts), <6GB VRAM, character/style LoRAs; covers setup, dataset prep, training params, and using the result in the anima-base workflow

Curated download URLs and target directories for every model the comfyui-mcp skills reference — checkpoints, VAEs, text encoders, LoRAs — organized by family (Flux, WAN, LTX, Qwen, Z-Image, SD15/SDXL). Use when downloading models with download_model / download_civitai_model, when a workflow fails with a missing-model error, or when setting up a new machine.

Authoring & publishing ComfyUI custom nodes to the Comfy Registry — node structure, pyproject.toml spec, comfy-cli publishing, and CI

Authoring ComfyUI v2 frontend extensions with @comfyorg/extension-api — defineNode/defineExtension/defineWidget, shell UI (sidebar tabs, commands, hotkeys), typed events, and handles. Use when writing or editing ComfyUI web-UI extension code (custom node JS, sidebar panels, widgets).

Build Qwen Image Edit workflows — model loading, conditioning, LoRAs, prompt patterns, and XY plot testing

Build Qwen Image 2512 text-to-image workflows — QwenImageIntegratedKSampler, separate component loading, lightning LoRAs, and fine-tuned model variants

Model family compatibility matrix — loaders, resolutions, samplers, CFG, VAE, ControlNet, and LoRA compatibility for SD 1.5, SDXL, Flux, SD3, and video models

ComfyUI prompt engineering knowledge — CLIP text encoding syntax, weight modifiers, model-specific prompting strategies, and best practices