H3 T2VA Draft

Text in, half a generation out, one image you can actually look at. H3 T2VA Draft is the short path into MiniMax H3's text-to-video-with-audio mode: you give it a model, a text encoder, both VAEs and a prompt, it builds the conditioning and the sigma schedule itself, runs part of the schedule, decodes the predicted first frame and stops there, waiting for you.

If you've used the older two-stage recovery trick on other video models - decode something ugly halfway, then re-run and pray the seed reproduces it - this is the honest version of that. The second half (H3 T2VA Continue) resumes the same generation with no new noise, rather than restarting the sampler and hoping.

Why you'd want a first-frame preview at all

H3 is a 33B omni-modal model that generates video and stereo audio together, up to 15 seconds at 2K/24fps - ComfyUI had day-zero support, and the launch-window reaction was unusually strong for a video release. It is also heavy, and the whole reason this pack exists is that the interesting failures are cheap to spot and expensive to re-roll: subject missing from the frame, an extra limb, the wrong room, the camera pointing somewhere useless. Live with a Turbo LoRA (the community's 4-step distillation recipe, strength 1, 6 total steps is the author's proven baseline) and a bad frame still costs you a full sampler pass plus a video and audio decode. Reviewing at step 3 of 6 costs half of that.

The inputs you actually set

Most of this node's sockets are loaders you already have wired:

• model, clip, video_vae, audio_vae - the H3 model (connect it after your Turbo LoRA loader), the H3 text encoder, and both H3 VAEs. The node checks these are a native MiniMaxH3 model and the matching H3 encoder/VAEs, so a Wan or LTX setup will get a clear error rather than nonsense.
• text - passed to H3 unchanged. No rewriting, no prompt magic: "Passed to native MiniMax H3 unchanged." The run also uses Core's basic guider path internally, so there's no CFG knob to fiddle with here; this is T2VA, text only.
• seed, width, height - seed defaults to 1234 (fixed, no auto-randomise), canvas defaults to 544 × 800, and both dimensions must be multiples of 32.
• duration_seconds - 1 to 15, and it snaps up to H3's 17k+5 frame grid. Five seconds is 124 frames at 24 fps, so about 5.17 seconds of video. Round numbers are notional here.
• total_steps / preview_steps - defaults 6 and 3. preview_steps "must be less than Total steps". Splitting 6/3 is the documented baseline; the schedule itself is Core's simple Euler path, so you're choosing a split point, not a sampler.

Outputs

preview_image is a single decoded frame at your canvas size - drop it into a Preview Image node or watch the panel the pack injects. draft_state is the important one: it carries the conditioning, the sigma schedule and the half-sampled AV latents, and it feeds H3 T2VA Continue and nothing else. report is a JSON string worth reading once - it prints the state ID, the frame count, the sigma you stopped at, and a flag that says out loud that the preview is not_a_guaranteed_final_frame.

Install

The pack ships no models and no pip dependencies; H3 weights, encoder, VAEs and Turbo LoRA are separate downloads from wherever you got them.

cd ComfyUI/custom_nodes
git clone https://github.com/ukr8b3g-cmyk/MiniMax-H3-T2VA-Draft-Continue
# restart ComfyUI

ComfyUI Manager works too - search "MiniMax H3 Draft Continue". You need a Core build with native H3 nodes, and a frontend recent enough for the review panel (ComfyUI 0.37.0 / Frontend 1.52.7 is the validated pair, on a 16 GB card).

Traps

Preview quality isn't final quality. You're looking at a step-3-of-6 x0 estimate. Composition is the signal; don't reject a clip over detail that hasn't been sampled yet.

Don't let it get stale. Anything upstream changing after Preview - prompt, seed, size, duration, step counts, a LoRA swap, even re-connecting the wire to the Continue node - invalidates the approval and the pack will refuse GO rather than silently continue a different run.

The two state types are not interchangeable. This node emits H3_DRAFT_STATE; the generic sampler pair emits H3_SAMPLER_DRAFT_STATE. Continuing one with the other's node errors on purpose.

Drafts don't survive a restart or a page reload, and saved workflows never carry approval. Templates you download always open needing a fresh Preview - including the author's own examples, which ship with GO off and an empty approval ID.

Watch the area. The author flags canvases over roughly 768 × 1344 as needing local memory verification, and the draft state has a hard 512 MiB budget once the AV latents are copied off the GPU.