LlamaServe-Doc H3 Omni Preset

This node doesn't generate anything. It formats. Give it a rough idea, a mode, a duration and some reference media, and it returns the two strings MiniMax-H3 wants: a system_prompt and a user_prompt, with your images, videos and audio baked in as official <Picture N> / <Video N> / <Audio N> labels. The pack's README is upfront that it's the legacy split-style node, kept alive so existing workflows keep working - the newer LlamaServe-Doc H3 Omni Generate merges this with generation into one box. If you're starting fresh, use that one. If you already have a Loader → Preset → Generate chain, this still does its job perfectly.

Why you'd reach for it

Two honest reasons. First, you've got an existing workflow wired around the pack's three-piece setup, and you don't want to rebuild it. Second, and this is the underrated one: because the Preset hands you the actual prompt strings, you can see exactly what your LLM is being asked to produce - paste them into any chat UI to eyeball the format before you spend GPU tokens on it. There's a real debugging value in seeing the system prompt the author embeds for each task mode.

How it works

Under the hood it calls the same build_h3_omni_preset routine the merged node uses, so the validation is identical. Mode must be one of the five official tasks - T2AV (text only), I2AV (exactly one image, first frame), L2AV (exactly one image, last frame), FL2AV (two images in order), or Ref2AV (one or more ordered references). duration is an integer 4–15 seconds, snapped to H3's 17*n+5 frame grid and written into the prompt to two decimal places. resolution is validated per mode - Ref2AV is locked to 16:9 or 9:16.

The Ref2AV path does the clever housekeeping: it converts Chinese aliases (图片1, 视频1, 音频1) into official tags for old workflows, and it only emits media your prompt actually mentions. Mention a label you didn't connect, or forget to mention one you did, and it errors out instead of silently sending a malformed request.

Inputs and outputs

Four required inputs, and the only one you'll agonize over is raw_prompt - for Ref2AV it must reference every media item you attach. mode, duration and resolution are all dropdowns/sliders with sensible defaults. The optional media input takes an LLAMASERVE_MEDIA bundle; if you're building the chain by hand, that's what carries your images and audio into the Generate node.

The three outputs - system_prompt, user_prompt, media - wire straight into LlamaServe-Doc Generate, which is the node that actually launches llama-server and streams the completion. Feed the system and user prompts to its two text inputs and the media bundle to its media input. That's the whole circuit; nothing here runs until Generate fires.

Install and gotchas

Same as the rest of the pack: ComfyUI Manager, search "LlamaServe-Doc", or git clone https://github.com/DocWorkBox/LlamaServe-Doc.git into ComfyUI/custom_nodes, then restart. No pip dependencies; the native llama-server binary downloads on first execution. You'll still need a Qwen GGUF in models/LLM - the demo workflow expects Qwen3.6-27B-H3-Prompt-Rewriter-Q4_K_M.gguf, with Qwen2.5-Omni + mmproj if you're doing multimodal.

The one trap worth naming: this is not an output node and it does zero inference. People wire it up, hit Queue, and wonder why nothing happens - because nothing is supposed to. The Preset only builds strings; the magic (and the model download, and the VRAM footprint) lives downstream in the Generate node. If you're new to the pack, skip the middleman entirely and use H3 Omni Generate, which does Preset + generation in one go and even emits a Director-ready groups output for MiniMaxH3Director.