Local ComfyUI image generation end-to-end: from a blank machine (download, install, start, fetch models, workflows) through queue/poll/download and chat embeds. Defaults to http://127.0.0.1:8188 (Flux.2 Klein + Qwen3-ready). Use when the user wants images and has Remedy + a chat provider — even if ComfyUI is not installed yet.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Local ComfyUI image generation end-to-end: from a blank machine (download, install, start, fetch models, workflows) through queue/poll/download and chat embeds. Defaults to http://127.0.0.1:8188 (Flux.2 Klein + Qwen3-ready). Use when the user wants images and has Remedy + a chat provider — even if ComfyUI is not installed yet.
Local image generation via ComfyUI
REST API, driven by Remedy’s built-in comfyui tool.
You can start with nothing installed. If the user asks for images and
ComfyUI is missing, follow From scratch bootstrap below (with their approval
for downloads / shell). Then generate with the native tool.
This skill pairs with Remedy Desktop (or CLI) connected to any LLM provider
(xAI, OpenAI, Ollama, …). The provider runs the agent; ComfyUI runs on the
user’s GPU/CPU for pixels.
Do notlist_dir the whole disk hunting for ComfyUI. Use action=locate /
status. Discovery is built in for every OS/user.
Agent decision tree (always start here)
Call comfyui / action=status (or locate).
API up → jump to Generate images.
Install found, API down → Start ComfyUI (use start_hint from the tool).
Nothing found → From scratch bootstrap (this is expected on a fresh PC).
After models are in place and the server is up → Generate images.
Never invent success. Never paste DSML/tool XML into chat. Prefer the comfyui
tool over raw curl (especially on Windows).
From scratch bootstrap (fresh install)
Goal: machine has no ComfyUI → user can generate a PNG from Remedy chat.
Ask once before large downloads (disk: tens of GB for full Flux.2 Klein;
smaller if using 4B). Confirm GPU type: NVIDIA (default), AMD, Intel, or CPU-only.
flux-2-klein-base-9b-fp8.safetensors (or a 4B Klein fp8 if VRAM-limited)
models/diffusion_models/
Text encoder
qwen_3_8b_fp8mixed.safetensors
models/text_encoders/
VAE
flux2-vae.safetensors
models/vae/
How to obtain them (agent guidance):
Prefer Hugging Face / official Comfy model docs for current Flux.2 Klein
and Qwen3 encoder packages. Filenames must match the table (or edit the
workflow / tool defaults if the user downloads alternate names).
ComfyUI Manager (if installed) or browser download → save into the folders above.
huggingface-cli download … is fine when the user has HF access and disk.
After downloads, restart ComfyUI so it rescans models.
VRAM guidance (agent should ask / recommend):
Hardware
Suggestion
≤8 GB VRAM
4B Klein (if available), 512×512, close other GPU apps
12–16 GB
9B fp8 often OK at 512–768
24 GB+
9B fp8 comfortable; higher res later
Rough disk: several GB to 20+ GB depending on UNET size — state this before download.
If the user only wants a smoke test and already has SD1.5/SDXL checkpoints, you
may queue a different API-format workflow they provide — but Remedy’s built-in
generate path expects the Klein filenames above unless you customize workflow JSON.
Phase D — Workflows (write / adapt)
Use the bundled starter (default)
Path (inside this skill):
scripts/workflows/txt2img_flux2_klein.json
Or let the tool build it: comfyui / action=generate with a prompt
(internally uses the same graph + injects the text prompt).
Write / edit a workflow
Preferred for custom graphs: open http://127.0.0.1:8188 → build in UI →
Save (API Format) → save .json under the project (e.g. assets/workflows/).
API format is a dict keyed by string node ids, each with class_type + inputs.
Links are [node_id, output_index] pairs — not the pretty UI format.
Edit with file_write / editor; keep filenames in loaders aligned with disk.
Queue via CLI helper or POST /prompt (see below). For chat UX, stick to
action=generate when the Klein graph is enough.
