comfyui_mcp

Diagnose and fix image quality issues in a ComfyUI workflow — face/body color mismatch, AI-looking skin, broken hands, anatomy artifacts. Use when the user shares a generated image and asks to improve it, when they say it "looks AI-generated", when hands/eyes/skin look off, or when iterating on a workflow's output quality. Includes diagnostic checklist, ranked fix recipes with concrete settings that have been validated, and escalation paths.

Craft a high-quality natural-language image prompt for Krea 2 (Krea 2 Turbo and Krea 2 Raw, the open-weights 12B text-to-image model with a Qwen3-VL text encoder, run locally in ComfyUI). Use when the user asks for a Krea 2 prompt, mentions Krea 2 / Krea2 / Krea Turbo by name, has a Krea 2 workflow open, or pastes a draft prompt to improve for Krea 2. Krea 2 is a dense-prose model with a built-in prompt enhancer — this skill IS that enhancer, so its output goes in with prompt_enhance turned OFF.

Find a LoRA (or other model — Checkpoint, ControlNet, Upscaler, Embedding, VAE) matching the user's request, prefer already-downloaded files, fall back to Civitai search when nothing local matches or the user explicitly asks. Walks the user through search → describe → confirm-download → inject into the open workflow. Auto-detects the open workflow's base model so cross-base mismatches (SDXL LoRA on Flux) are never suggested. Use when the user asks "find me a LoRA for X" / "is there a local LoRA I can use" / "search civitai for X" / "I need a [thing] LoRA" / "what controlnet do I have for openpose" / "find a Flux checkpoint."

Craft a high-quality natural-language image prompt for Flux 1 (Dev / Pro / Schnell), Flux 2 (Pro / Max / Flex / Dev / Klein), and Flux NSFW fine-tunes (Flux-Uncensored, Lewd Flux, etc.). Use when the user asks for a Flux prompt, asks to enhance a basic idea into a Flux prompt, mentions any Flux variant by name, or pastes a draft prompt and asks for it to be improved for Flux. Flux does NOT support negative prompts — the skill handles that.

Craft a high-quality video prompt for the LTX family from Lightricks — LTX Video 2B/13B, LTX-2 (audio+video, 4K, 50fps), LTX-2.3 (current SOTA, 22B). Use when the user asks for an LTX prompt, mentions LTX / LTX-Video / LTX-2 / LTX-2.3 / Lightricks / LTX Studio by name, wants real-time video generation, wants synchronized audio-video, or wants to use LTX's camera control LoRAs. The skill follows Lightricks' official prompt structure and supports negative prompts.

Craft a high-quality image prompt for Qwen Image (Alibaba) — Qwen Image 1.0, Qwen Image 2.0 (Feb 2026, current SOTA), and Qwen Image Edit for image-to-image editing. Use when the user asks for a Qwen prompt, mentions Qwen Image / Qwen-Image / Qwen3-VL by name, asks to enhance a description for Qwen, or wants to edit an existing image with Qwen Edit. Qwen supports negative prompts. Killer features are professional-grade text rendering (including Chinese) and unified generation + editing in one model.

Craft a high-quality video prompt for the Wan family from Alibaba's Tongyi Wanxiang Lab — Wan 2.1, Wan 2.2 (T2V / I2V / Animate / Fun-Control), Wan 2.5, Wan 2.6 (shot-block cinematographer-style prompting), and Wan 2.7 (Thinking Mode, multi-reference). Use when the user asks for a Wan prompt, mentions Wan / WAN / Tongyi Wanxiang / 通义万相 by name, wants to convert a still into video (I2V), wants text-to-video generation, or wants help with cinematic multi-shot video prompts. Wan supports negative prompts and the skill handles them.

Craft a high-quality image prompt for the Z-Image family from Alibaba's Tongyi-MAI Lab — Z-Image (base, Jan 2026, supports negative prompts), Z-Image-Turbo (8-step distilled, ignores negative prompts), Z-Image-Omni (unified gen+edit), and Z-Image-Edit. Use when the user asks for a Z-Image prompt, mentions Z-Image / Tongyi / Tongyi-MAI by name, asks to enhance a description for Z-Image, or wants to edit an existing image with Z-Image-Edit. The skill handles negative-prompt support correctly per variant.

Craft a danbooru-tag image prompt for Illustrious XL, Pony Diffusion V6 XL, or other SDXL fine-tunes trained on danbooru tags. Use when the user asks for a Pony, Illustrious, or SDXL danbooru-style prompt. Returns positive AND negative prompts.

Decide HOW to build a mask before running an inpaint pass. Use when the user wants to inpaint, mask, or regenerate a region of an image, OR when an inpaint failed (mask seam visible, hallucinated body parts, two-halves anatomy). The wrong mask source guarantees a wrong inpaint; this skill picks the right tool for the region BEFORE wiring up nodes.

Extract, edit, and synthesize OpenPose skeletons for ControlNet conditioning. Use when the user wants a unique/custom pose, wants to modify a pose extracted from a reference image, or wants to escape the "AI default poses" attractor by feeding ControlNet a hand-tuned skeleton. Covers the extract → modify → render → ControlNet pipeline with concrete coordinates and PIL recipes.