MiniMax H3 Progressive Long Video / 渐进双采长视频 (EXP/T8)

Every long-video workflow in ComfyUI eventually becomes the same puzzle: generate a window, stitch it to the last one, keep the audio from drifting, and don't lose an hour of work when a segment fails. This node does all of that internally. You give it a configured model pair, a sampler, one full sigma table, CLIP and the two VAEs, plus a duration - it renders the video in segments, splits each segment's schedule into a cheap low-res first pass and a full-res finish, carries the accepted picture forward as context, and writes the MP4 itself.

It is an output node. You don't add a SaveVideo afterwards.

Defaults that tell you the author's mood

filename_prefix defaults to H3_Progressive_UNREVIEWED, and the code raises an error if the delivery report ever claims human qualification. That is not decoration - it's the pack's way of saying this route has real mechanical verification (two 8-second segments with real weights, strict video and audio decode, human review) but that a full combination matrix and quality sign-off are still open. Take it seriously. Run it, watch it, decide for yourself.

The inputs worth touching

Wire in model, clip, video_vae, audio_vae, sampler and sigmas from the Progressive Dual Setup node - those six define the run and they must come from the same contract.

• chain_id - the identity of this run. Same id plus identical inputs means "verify and resume"; change the model, LoRA, prompt, source material, duration or resolution and you should change the id rather than drag a failed segment's cache into a new experiment.
• total_duration_seconds, width, height - target output, dimensions snapped to 32.
• render_window_frames (124–362, on a 17n+5 grid) - the internal generation window. Here's the counterintuitive bit: even a 3-second output renders at least a 124-frame window and trims down. You cannot ask it for 73.
• low_evaluations - how many steps of the full 8-step table run cheap. 4 gives you 4+4, 6 gives you 6+2. The node requires at least one HIGH step.
• low_scale - the LOW spatial fraction. At 0.5 with an 896×448 target, the first pass is 448×224.
• context_frames - 22 or 39 frames of accepted-picture context carried into the next segment.
• upscaler_model - pick your existing learned 3D latent upscaler. Nothing is auto-downloaded.
• global_prompt / segment_prompts_json - the standing description of the scene versus per-segment events. One-off dialogue belongs in the segment events, not the global prompt you resend every window.
• resume_existing - true re-verifies real models, sources, settings and on-disk receipts before reusing stages. false refuses to touch an existing chain inside an OS lock and never deletes your files.
• audio_mode, drive_audio, final_audio, audio_seam_policy (a 5 ms cosine bridge by default) - the audio side. final_audio is preferred as the output track; being the output track doesn't mean it drove generation.
• first_frame - a single first frame for segment 0. That's the extent of the image conditioning.
• model_hires (optional) plus the optional prompt_relay_plan, and the eav_* group, which applies the pack's EAV gain internally. Don't feed in a MODEL that already carries an external EAV wrapper, and don't run the TST node and the internal EAV settings as two competing configurations of the same branch.

Outputs: video (a VIDEO object), video_path (the rendered file), and report_json (per-stage record - read it when a segment decides to redo itself).

Timing math you'll otherwise get wrong

An 8-second, 24 fps output is 192 frames, and with a 124-frame window the seam lands around frame 124, about 5.17 seconds in. It is not two 4-second halves. Short clips pass; that says nothing about a 24- or 32-second run.

Scope, and the honest gaps

T2VA and single-first-frame I2VA only. Not multi-reference, not end-frame, not reference-audio editing. For those, use the pack's older dual-MODEL in-node long-video route - this node doesn't replace it, and the old workflow's defaults didn't change.

OOM is still OOM: lower resolution, frames and reference count, run one H3 job at a time, and remember reserve_vram_mib checks the boundary rather than capping your card. Audio weirdness gets the same checklist as everywhere else in this pack - sampler, scheduler, steps, video/audio shifts, LoRA, before you blame the node. And installing is the usual: Manager → MiniMax H3 Audio T8 → full restart, no pip extras, plus ffmpeg on PATH so the delivery step can write the file.