loading…
loading…
<rag>…</rag> tag inside your message. The server injects top-K snippets as context before the model answers, and citations appear above the reply.<rag collection="name">…</rag> to restrict to one collection.Model picker (top bar) lists every discovered chat model plus the configured image/video/audio/3D providers. Click a media provider, type a prompt, and the response is generated via ComfyUI on GPU 0.
Temperature controls randomness. Low for code/facts; high for brainstorming.
Top P limits the vocabulary pool. Usually leave at 0.9.
Frequency/Presence penalty discourage repetition or encourage new topics. Usually 0.
Stop sequences kill the stream on specific tokens.
Seed (integer) makes outputs reproducible when combined with the same prompt + params.
Select an image provider from the dropdown (FLUX.2-dev, Z-Image Turbo, Chroma1-HD, FLUX.1-schnell). Your message is the prompt.
Steps: schnell/turbo need only 4; dev/HD models want 20–30. More ≠ better after ~40.
CFG Scale: how strictly to follow the prompt. 1 for schnell, 3–5 for dev, 7+ for ultra-literal.
Seed: reproduce an exact image by reusing its seed (shown under every output).
Select FLUX.1-Kontext, attach a reference image using the paperclip icon, and type the edit instruction (e.g. "make the sky purple", "remove the car"). The attached image is passed as the edit target.
Select HunyuanVideo 1.5 or LTX-Video. Type the prompt. Expect 2–5 min per clip depending on frames/resolution. HunyuanVideo native fps is 16; LTX is 24.
Select Stable Audio Open, type a description of the sound (e.g. "heavy rain on a tin roof"). Duration is capped at 47s. Output plays inline.
Select Hunyuan3D 2.1, attach a clear front-facing image of an object, submit. Output is a downloadable .glb mesh.
/image <prompt> · /video <prompt> · /audio <prompt> · /edit <instruction> · /3d force-route to the named pipeline regardless of currently-selected model.
Click the status chip in the top bar, then use the "GPU N" dropdown to swap which chat model is loaded on that GPU. Other instances on the same GPU are stopped automatically. Models need 1–5 min to warm up before they appear in the chat dropdown.
Enter = send · Shift+Enter = newline · Esc = close panels
Every conversation can be exported as Markdown via the Export button in the top bar. Files include model, date, and the full transcript.