A set of OpenWebUI Functions (pipe models) that turn a local Ollama + ComfyUI stack into an auto-routing assistant plus standalone image/video generators. Install them from Workspace → Functions — they run inside the stock OpenWebUI image, no fork or source patch required.
| Function | Model | What it does |
|---|---|---|
auto_assistant |
🪄 Assistant | One model that decides: chat (with vision), a Krea 2 image (create or edit), a Wan 2.2 video, or animate a still. Routes by intent and manages GPU VRAM around every job. |
image_krea |
Image | Krea 2 Turbo text-to-image + Qwen-Image-Edit instruction editing, local prompt-enhance, vision-QA correction. |
flux_image |
— | FLUX.1-dev text-to-image pipe. |
animate_scail |
Animate | SCAIL motion-transfer: drive a still with a motion clip. |
- OpenWebUI (official image) with Ollama (
localhost:11434) and ComfyUI (localhost:8188) reachable. - Ollama: a chat model (e.g.
dolphin-venice:24b) + a vision model (gemma4:31b) for image QA. - ComfyUI with the referenced checkpoints/GGUFs (Krea 2 Turbo, Wan 2.2 A14B, Qwen-Image-Edit, …).
- Intent routing with question/small-talk guards — a question about an image isn't turned into an edit.
- Wan 2.2 A14B video incl. multi-shot sequences with per-shot frame QA.
- Vision-QA verification of every generation against the request, with corrections.
- VRAM choreography — a generation lock, real VRAM-release polling, idle-gated ComfyUI unloads.
- Native status line — a live elapsed-time ticker collapsing to
Generated in 1m 05s · Krea 2 · 1024×1024 · 8 steps.