[PR #3175] [MERGED] Add OpenH3-IR to custom-node-list.json #3075

New issue

Closed

opened 2026-10-28 21:14:58 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-28 21:14:58 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/Comfy-Org/ComfyUI-Manager/pull/3175
Author: @ruashots
Created: 8/16/2026
Status: ✅ Merged
Merged: 9/22/2026
Merged by: @ltdrdata

Base: main ← Head: add-openh3-ir

📝 Commits (4)

045d3d5 Add OpenH3-IR to custom-node-list.json
5db01ec Say what the nodes need: this repo's service and a model you already run
864899a Point at the nodes' own repository
8809373 Nothing to start: the pack brings the compiler with it

📊 Changes

1 file changed (+11 additions, -0 deletions)

View changed files

📝 custom-node-list.json (+11 -0)

📄 Description

Adds the ComfyUI node pack for OpenH3-IR.

MiniMax published the format for H3's Context-IR document, the long structured prompt the model
actually wants, along with finished examples of it. What they kept on their own servers is the part
that writes it. OpenH3-IR is an open take on that part. It is not a prompt enhancer. It writes the
document to MiniMax's published format, with the named sections in a fixed order, numbered picture
labels, a length on H3's frame grid, per-shot cut times, and sound described separately. None of
that is left to the model to get right. The structure is computed rather than asked for, the
finished document is validated against the format, and one that does not hold is refused and
written again.

The four nodes this repository ships. They are how OpenH3-IR is used from ComfyUI, and they
bring the compiler with them, so the pack is the whole thing rather than a front end for something
you install separately. Main takes the prompt and hands the render the model, conditioning, latent and
both VAEs, so the graph needs no loader boxes and no H3 conditioning node. Media is one tray for
pictures, clips and sounds, each carrying the name the prompt refers to it by. Setup holds the
language model address and the five H3 files. Director is optional reusable direction.

Repo: https://github.com/ruashots/ComfyUI-OpenH3-IR (Apache-2.0)

What a reviewer should know.

Dependencies. requirements.txt declares one package, open-h3-ir, published on PyPI under
Apache-2.0. Manager's pip step installs it. Nothing is vendored, and no copy of the compiler lives
in this repository, so either half can be updated on its own.

Nothing to start. The compiler runs inside ComfyUI's own Python. There is no service to launch and
no port to pick. Pointing the graph at an OpenH3-IR instance on another machine over HTTP stays
available and is not the normal path.

The one thing a user needs is an OpenAI-compatible model endpoint they run or have access to:
vLLM, llama.cpp server, LM Studio, Ollama, or a hosted one. No cloud account is required, and the
compiler wants no GPU of its own. If the endpoint is missing or down, the node says which and names
the field to fill in.

The import is lazy. The nodes do not import the compiler while ComfyUI is loading them. A missing,
half-installed or broken open-h3-ir costs the compile rather than the pack: the nodes still appear
on the menu, and the failure is reported when a graph tries to compile.

Install shape. The repository root is the pack. A git-clone into custom_nodes loads as-is, with
no subfolder and no re-export.

Checks. 519 tests in this pack and 990 in the compiler, all of them with no model, no GPU and no
network. MiniMax's own published examples are in the compiler's test set and have to pass clean,
because a check that fires on their own artifact is a wrong check. This pack also carries a harness
that plants 113 deliberate defects one at a time and requires the guard for each one to fail.

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/Comfy-Org/ComfyUI-Manager/pull/3175 **Author:** [@ruashots](https://github.com/ruashots) **Created:** 8/16/2026 **Status:** ✅ Merged **Merged:** 9/22/2026 **Merged by:** [@ltdrdata](https://github.com/ltdrdata) **Base:** `main` ← **Head:** `add-openh3-ir` --- ### 📝 Commits (4) - [`045d3d5`](https://github.com/Comfy-Org/ComfyUI-Manager/commit/045d3d585d6a862a0d70d331dccaf63cf28e6b4b) Add OpenH3-IR to custom-node-list.json - [`5db01ec`](https://github.com/Comfy-Org/ComfyUI-Manager/commit/5db01ece5b188f6a60dfb85002dc73363afba9a3) Say what the nodes need: this repo's service and a model you already run - [`864899a`](https://github.com/Comfy-Org/ComfyUI-Manager/commit/864899a6a87292c3ae2b43cee5d85225e5e1f5eb) Point at the nodes' own repository - [`8809373`](https://github.com/Comfy-Org/ComfyUI-Manager/commit/8809373bf90ed86484f76ab58cbc748e533eebde) Nothing to start: the pack brings the compiler with it ### 📊 Changes **1 file changed** (+11 additions, -0 deletions) <details> <summary>View changed files</summary> 📝 `custom-node-list.json` (+11 -0) </details> ### 📄 Description Adds the ComfyUI node pack for OpenH3-IR. MiniMax published the format for H3's Context-IR document, the long structured prompt the model actually wants, along with finished examples of it. What they kept on their own servers is the part that writes it. OpenH3-IR is an open take on that part. It is not a prompt enhancer. It writes the document to MiniMax's published format, with the named sections in a fixed order, numbered picture labels, a length on H3's frame grid, per-shot cut times, and sound described separately. None of that is left to the model to get right. The structure is computed rather than asked for, the finished document is validated against the format, and one that does not hold is refused and written again. **The four nodes this repository ships.** They are how OpenH3-IR is used from ComfyUI, and they bring the compiler with them, so the pack is the whole thing rather than a front end for something you install separately. Main takes the prompt and hands the render the model, conditioning, latent and both VAEs, so the graph needs no loader boxes and no H3 conditioning node. Media is one tray for pictures, clips and sounds, each carrying the name the prompt refers to it by. Setup holds the language model address and the five H3 files. Director is optional reusable direction. Repo: https://github.com/ruashots/ComfyUI-OpenH3-IR (Apache-2.0) **What a reviewer should know.** *Dependencies.* `requirements.txt` declares one package, `open-h3-ir`, published on PyPI under Apache-2.0. Manager's pip step installs it. Nothing is vendored, and no copy of the compiler lives in this repository, so either half can be updated on its own. *Nothing to start.* The compiler runs inside ComfyUI's own Python. There is no service to launch and no port to pick. Pointing the graph at an OpenH3-IR instance on another machine over HTTP stays available and is not the normal path. *The one thing a user needs* is an OpenAI-compatible model endpoint they run or have access to: vLLM, llama.cpp server, LM Studio, Ollama, or a hosted one. No cloud account is required, and the compiler wants no GPU of its own. If the endpoint is missing or down, the node says which and names the field to fill in. *The import is lazy.* The nodes do not import the compiler while ComfyUI is loading them. A missing, half-installed or broken `open-h3-ir` costs the compile rather than the pack: the nodes still appear on the menu, and the failure is reported when a graph tries to compile. *Install shape.* The repository root is the pack. A git-clone into `custom_nodes` loads as-is, with no subfolder and no re-export. **Checks.** 519 tests in this pack and 990 in the compiler, all of them with no model, no GPU and no network. MiniMax's own published examples are in the compiler's test set and have to pass clean, because a check that fires on their own artifact is a wrong check. This pack also carries a harness that plants 113 deliberate defects one at a time and requires the guard for each one to fail. --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-28 21:14:58 +00:00