MiniMax H3 ONE STUDIO (TJ)

MiniMax H3 ONE STUDIO (TJ) is this pack's most ambitious node and its least finished - a full UI for MiniMax H3, the new open-weight video+audio model from Comfy Org. H3 is brand new - basically all community discussion landed in mid-2026 - so treat everything here as living on the frontier; the node is explicitly experimental and earns the label.

The interesting part isn't the model, though - it's the architecture. H3 only accepts frame counts on a 17k+5 grid, and one long generation blows past VRAM. So this node renders a long brief as sequential clips, saves each to disk, then stitches them into one file with ffmpeg. Each clip is its own queue submission, and VRAM is freed explicitly between clips and at run end - ComfyUI would otherwise hold the model resident forever.

Modes and continuity

You get three entry modes: Text only, First/Last Frame (start plus optional end keyframe), and Reference (up to nine reference images addressed as <Picture 1>; it needs its own UNET set in Settings). One prompt equals one clip, so total length is clips × clip length - add prompts or split a long brief on [Shot N] markers or blank lines. The style preamble and sound tail lift into a shared header/footer every clip carries, so a longer piece keeps a consistent look.

Then there's the Continuity menu - the part nobody else builds. Order is None → One-Take (latent) → Reference → Last Frame Chain, and One-Take is the default. It feeds the previous clip's sampled latent tail straight into the next clip's head via disk checkpoints (TJ_H3_SaveLatentCheckpoint / LoadLatentCheckpoint, from the author's TJ_NODE pack) - no VAE round-trip, overlap preserved to float32 noise. The overlap is a fixed 39 frames (1.625s), not adjustable: it has to match what's baked into the latents or the stitch comes out wrong. Last Frame Chain is the simpler sibling - the previous clip's final frame becomes the next clip's first frame, which forces FL2VA for the continued clips.

Inputs, outputs, and what they're for

Like the rest of the pack, the schema is minimal: no required inputs, two optional ones - prompt_override (prepended before the internal prompt) and pipe (the optional TJ_NODE PromptDB pipe that overrides prompt/seed/steps/CFG/sampler/scheduler at generation, without touching your UI). The two outputs matter more:

• last_frame (IMAGE) - the final frame of the last generated clip. Wire it to a preview or use it as a follow-up's first frame.
• video_path (STRING) - the filesystem path of the saved video, either the individual clip or the stitched result, ready for a file-watcher node or anything that takes a path.

Models - the setup tax

MiniMax H3 needs a lot of files from Comfy-Org/MiniMax-H3: two UNETs (one for First/Last, one for Reference), a Qwen3-VL text encoder (loaded with type=minimax), and two VAEs - video and audio, both required. Here's the trap: the audio slot must hold an minimax_h3_audio_vae_* file. Point it at another audio VAE - LTX's, say - and decode dies with a tensor-size mismatch. A mode whose UNET isn't set just won't let you in - the pills grey out and the top bar says what's missing.

Optional packs bolt on speed (SolAttn is the default accel, plus Spectrum, Turbo LoRA and a step-reuse cache) and everything degrades gracefully - a missing pack just switches its feature off. Turbo is FL2VA-only, so its absence in Reference mode is expected, and enabling it force-disables the cache because a 4-step schedule never reaches the cache's reuse threshold. Stitching needs ffmpeg; if it fails, pip install imageio-ffmpeg usually fixes it.

Install and the one gotcha that bites everyone

cd ComfyUI/custom_nodes
git clone https://github.com/designloves2/ComfyUI-TJ_NODE_STUDIO_ONE.git

or find it in ComfyUI Manager and restart. The pack ships install_requirements.bat / .sh to pull the optional node dependencies for you; One-Take continuity specifically wants ComfyUI-TJ_NODE from the same author.

And the gotcha: generation runs from the node's own ▶ Generate button, not ComfyUI's RUN. RUN just re-outputs the last result from the output slots. You'll hit this within your first five minutes. Stitched long videos also aren't seamless - the README admits it - so a visible cut is the relay design, not you. Clips feeling unrelated is the fixable one: switch Continuity to Last Frame Chain and keep shared style/sound text in the COMMON prompt. Given how new H3 is, set expectations accordingly - this is a tool for people who like being early.