Nodes/ComfyUI-OpenH3-IR/OpenH3-IR Setup
ComfyUI Node

OpenH3-IR Setup

Which language model writes the brief, which five H3 files to load, and where the compiler runs. Every H3 graph needs one of these: the compile node loads what you pick here.

By ruashots·Created 3 days ago·Updated 2 days ago· 4
OpenH3-IR Setup
    • setup
    ◄server►
    ◄reference_model▾►
    ◄frames_model▾►
    ◄text_encoder▾►
    ◄video_vae▾►
    ◄audio_vae▾►
    ◄weight_dtypedefault►
    ◄timeout_s600►
    ◄llm_url►
    ◄llm_model►
    CategoryOpenH3-IR

    Inputs (10)

    NameTypeDefaultDescription
    serverSTRINGLeave empty to compile in ComfyUI itself, which needs nothing started. To compile on another machine, put the address an OpenH3-IR service is listening on there instead, for example http://127.0.0.1:8420. That service uses its own language model, so the two fields below are then not used.
    reference_modelCOMBOH3's checkpoint for reference and text jobs. Both formats are in this list: pick a .gguf and it loads through Unet Loader (GGUF), pick a .safetensors and it loads like Load Diffusion Model does.
    frames_modelCOMBOH3's checkpoint for first and last frame jobs. The compile node uses this one or the ref2va model depending on which slots you filled, and says which in its report. Both formats are in this list.
    text_encoderCOMBOThe Qwen3-VL encoder H3 was trained against, the same file a Load CLIP node takes. Both formats are in this list, chosen independently of the checkpoint: a GGUF clip works with safetensors weights and the other way round.
    video_vaeCOMBOH3's video VAE, used for the decode as well.
    audio_vaeCOMBOH3's audio VAE, a different file from the video VAE. Needed even for a silent piece, because H3 writes picture and sound together.
    weight_dtypeCOMBOdefaultThe same setting a UNET loader has. Leave alone unless you are short of VRAM. It does not apply to a GGUF checkpoint, which carries its own quantisation, and the report says when it was ignored.
    timeout_sINT60010–3600Writing a brief is one call to your language model, so this is as slow as that model is.
    llm_urloptSTRINGThe OpenAI-compatible endpoint the brief is written with, in full and ending in /v1, for example http://192.168.1.20:8000/v1. vLLM, llama.cpp's server, LM Studio, Ollama or a hosted API all work. It has to be able to read pictures: every reference in the tray is described through it.
    llm_modeloptSTRINGWhich model on that endpoint, by the id it serves. Leave empty on an endpoint that serves one model. On one that serves several this will not guess, because no model list says which model can see.

    Outputs (1)

    NameTypeDescription
    setupH3IR_SETUP—