News

The node has a walkthrough now: 34 minutes covering every documented MiniMax H3 feature, in the order you meet them.

• The timeline. Tracks for the picture, the camera and the sound; the playhead, splitting a block, and why clip length snaps to 5 + 17 frames.
• Entities out of a file. A person, a costume, a prop, a place, a style, a motion, a voice — each becomes a card with a number you paste into any block, and it stays the same across shots.
• Dialogue. Describe a voice, pick who speaks, in what language, on or off screen.
• References. Storyboard, first and last frame, keyframe, and how much of a file survives — including carrying a face from one photo onto a person in another.
• One full example built from scratch and generated at the end.

Watch it: MiniMax Director — the full walkthrough. Dubbed: Русский, 中文.

It is already on your GPU. Start the MiniMax H3 bundle and open the MiniMaxDirector workflow — the bundle loads the newest release from GitHub. On your own machine, install minimax-director from the ComfyUI registry, or clone github.com/imbutus/ComfyUI-MiniMaxDirector into custom_nodes/.

Three models joined the LLMs page and two left it.

• huihui-ai/Huihui-Ornith-1.5-9B-abliterated — small, cheap, and tuned for code. A 9B with image input and 262k context, abliterated. It is also the first model here carrying two labels at once: it is the budget tier and a coding model, so you will find it under either heading.
• huihui-ai/Huihui-Ornith-1.5-35B-A3B-abliterated — the bigger Ornith. Same family, a 35B mixture-of-experts, image input, 262k context, abliterated. It takes over from Ornith 1.0.
• huihui-ai/Huihui-CyberStrike-OffSec-35B-abliterated — a security specialist, in a category of its own. A 35B tuned on penetration testing and red-team work, image input, 262k context, abliterated. Models like it now carry an offsec label instead of sitting among the general-purpose ones.
• Two models retired. The Qwen3.6 35B-A3B and the Ornith 1.0 35B are no longer offered: Ornith 1.5 covers what they did, and the Qwen3.8 27B stays the general pick. Sessions you already ran on them keep showing in your history.
• Labels stack now. A model is no longer limited to a single badge, which is what lets the 9B appear as both cheap and coding.

Open the LLMs page to start any of them.

The node I released in August has had 264 commits since, and 0.16.0 collects them. The idea is unchanged — you lay a film out on a timeline instead of cramming it into one prompt box — but far more of what H3 actually reads now comes out of it.

• Everyone and everything gets a card. A person, a prop, a costume, a place, a look: one card each, holding the picture it came from, its description and its voice. It is the only place a file is described, so nothing is repeated on a block.
• The voices come out of the same pass as the picture. Write the line, pick the face, and it compiles in the exact form H3 was trained on — including a line that crosses a cut, or one cut off by the end of the clip.
• A face can be carried onto somebody else. A card can also take its motion from a video and its timbre from a recording, and each of those is cited inside the subject it belongs to rather than left floating.
• Every file the clip carries sits in one list. Drag one onto a track to place it, or leave it clip-wide for a look that has to hold throughout. A file that has gone missing says so, and the run is refused instead of quietly producing the wrong thing.
• A live linter checks the rules the guides state outright: clip length, reference clips outside 2–15 seconds, a marker that cannot do what it is being asked to do.
• The node is as tall as whatever you have open, and the whole piece exports as one JSON.

Full notes: the 0.16.0 release.

It is already on your GPU. Start the MiniMax H3 bundle and open Docs → Media → MiniMax H3 → the MiniMaxDirector workflow. The bundle loads the newest release from GitHub, so there is nothing for you to update.

On your own machine, clone github.com/imbutus/ComfyUI-MiniMaxDirector into custom_nodes/, or install minimax-director from the ComfyUI registry once 0.16.0 clears review there. Needs ComfyUI 0.31.0 or newer. MIT licensed.

The newest Qwen generation is on the LLMs page: huihui-ai/Huihui-Qwen3.8-27B-abliterated, a dense 27B in full precision, abliterated.

• It takes longer to answer, and the answers are smarter for it. Reasoning runs at maximum effort by default, so the first tokens come noticeably slower than on the smaller models. The wait is the whole point: it works the problem through instead of replying fast. If you want a quick back-and-forth, one of the smaller models is the better pick.
• It reads images. Attach a picture and ask about it, in the same conversation as everything else.
• 262k context. Long enough for a whole codebase or a stack of documents in one session.
• Full precision, nothing thrown away. The weights are the ones the authors published, which is why it needs a large card and about 10 minutes before the first message.

$2.40/hr. Open LLMs and start it.

I wrote a ComfyUI node for MiniMax H3 and released it as open source: MiniMaxDirector. Instead of packing an entire film into one prompt box, you lay it out on a timeline — what happens on screen, how the camera moves, what is heard — and it compiles that into the single structured prompt H3 actually reads.

What it does for you:

• Three tracks. Shots, camera and audio, each segment with its own text and its own span.
• Cut times are computed, so the model is told exactly when each shot begins.
• Legal lengths only. The clip is snapped to a duration H3 accepts, so a run cannot fail on arithmetic.
• Attach a file to a segment and it becomes a numbered reference in the prompt automatically — no tokens to type by hand.

It is already installed on your GPU. Start the MiniMax H3 bundle and open Docs → Media → MiniMax H3 → the MiniMaxDirector workflow; the graph is ready to run, nothing to install.

On your own machine, it is on the ComfyUI registry as minimax-director — searchable in the ComfyUI Manager, or comfy node install minimax-director. Source and issues: github.com/imbutus/ComfyUI-MiniMaxDirector. MIT licensed.

The MiniMax H3 bundle now has a walkthrough on the Docs page, in English, Russian and Chinese, with subtitles.

It goes through a first clip end to end:

• Where the workflows are and which one to open first.
• Writing sound into the prompt — what to describe, and what the model does with it.
• Legal clip lengths — why an arbitrary duration is snapped, and which numbers land on whole seconds.
• Attaching references so a subject stays the same from one shot to the next.

Watch it on YouTube, or open Docs → Media → MiniMax H3.

A new video bundle is on the Media page: MiniMax H3, an omni-modal model that generates picture and sound in one pass — the audio is not dubbed on afterwards, it comes out of the same generation as the video.

What is worth knowing before your first clip:

• Sound is part of the model. Describe what is heard the same way you describe what is seen, and it is generated with the shot. Stereo, no separate voice step.
• Reference images, audio and video. Attach them and point at them from the prompt; the model uses them to keep a subject or a look consistent across shots.
• Clip length is not free. H3 accepts lengths on a fixed lattice at 24 fps, so the workflow snaps whatever you ask for to the nearest legal value. 8, 25 and 42 seconds are the whole-second ones.
• One prompt, many shots. The model reads a shot list with cut times, so a single generation can contain several shots rather than one continuous take.

Open Media → Video → MiniMax H3 to start it.

The four voice bundles now have one clear job each, one shared set of workflow names, and they start faster.

• Fish Audio S2 — the dubbing pick. Text-to-speech and voice cloning across 80+ languages, with voice-to-SRT and SRT-to-voice that keep the original timing.
• CosyVoice 3 — the change-voice pick. Native voice conversion keeps the original words, pauses and delivery and swaps only the timbre, with no transcription step in between.
• Qwen3-TTS — the all-Qwen bundle. Three-second voice cloning plus voice design, and the only bundle that ships Qwen3-ASR as a second transcription engine.
• WhisperX is now the transcription engine everywhere. It is the more accurate one in real use, and dropping the second engine from the other three bundles removed about 3.7 GB from each pod, so voice pods are ready sooner.
• Two workflows you did not have before. common-align-script-to-srt is in every voice bundle now — paste a script with a blank line between sections and it aligns them to your narration as an SRT. CosyVoice 3 also gained a proper cosyvoice3-voice-clone.

Some workflow names changed in the _examples folder, so look for the new ones: *-tts-clone is now *-voice-clone, *-audio-to-srt is now common-audio-to-srt, and qwen3tts-change-voice is now qwen3tts-redub-voice — it transcribes and re-speaks the words rather than converting the voice, so use CosyVoice 3 when you want the delivery kept.

The Docs page now lists every bundle with its models and every workflow it ships, each with the same notes you see on the ComfyUI canvas.

The Sulphur-2 bundle now has a full walkthrough on the Docs page — just under eight minutes, in English, Russian and Chinese, with subtitles.

It covers the LTX Director workflow end to end:

• Image → video — make a person in a photo say the text you give it.
• Image + audio → video — feed a voice clip you prepared in Fish Audio S2 and have the person in the picture speak that exact recording.
• Extending a clip — put an existing video at the start or the end of the timeline and let the model generate the missing part. Match the framerate of your source, or the result drifts.
• fp8 vs bf16 — these are different models, not quality presets. fp8 is the compressed one: lighter on resources, weaker results.

Also covered: why to test on a short time span before committing to a full render, why splitting prompts into short intervals generates more accurately, and how the wrong duration makes the model hallucinate. The voice itself can be replaced afterwards with CosyVoice3.

Open Docs → Media → Video tutorial to watch it.

SCAIL-2 has been through a full round of testing and is out of beta. Both of its jobs work end to end, and the Beta chip is gone from the bundle card.

• Character animation — drive your own character with the motion from any clip. The background is generated fresh; the driving video's scene is not kept.
• Character replacement — swap one person in an existing video and keep the original scene, background and everyone else untouched.
• Two characters at once — one reference photo holding both of them, animated together from a clip with two moving subjects.

Every workflow now opens with step-by-step notes inside ComfyUI: which inputs to fill, how to point the tracker at one person when the clip has several, what to check in the mask previews before committing to a long render, and how to go past the default five seconds.

Two things worth knowing before you start. The render is silent — SCAIL-2 generates picture only, so add sound afterwards in a video editor. And the reference image has to be a real photograph; for two characters that means one picture of both of them together, not two photos joined side by side.

A video tutorial is coming later.