MiniMax H3 Chain Orchestrator / 总时长自动分段 (EXP/T8)

MiniMax H3's generation window is fixed - it produces specific grid lengths, not an arbitrary "make me 73 seconds." So making a long video means running a chain of segments, and somebody has to decide how many segments, how long each is, and where to pick back up after a crash. That somebody is MiniMaxH3LongVideoOrchestratorT8, and it's the most hands-off node in the Long Video cluster: you give it a total duration and the render window, and it turns that into a full H3 timeline.

The orchestrator's job, in three pieces: convert the total duration into fixed-window segments on H3's 17n+5 frame grid, pick the first segment that hasn't been accepted yet (so an interrupted chain resumes instead of restarting), and - the detail that tells you this author thinks - block downstream sampling once the final manifest is complete, so you can't accidentally generate an extra segment past the end of your video. That last bit is the difference between a planner and a thing that loops forever.

Inputs you'll actually touch

• total_duration_seconds - the whole video's length, up to 3600s. This is the number that drives everything.
• render_window_frames - 124 by default, and the tooltip is worth quoting: keep 124 for the current bounded-memory baseline; larger windows need separate VRAM validation. Don't crank this because "362 is more."
• context_frames - the head-context overlap between segments (5/22/39).
• global_prompt and segment_prompts_json - a base prompt for every segment, plus an optional JSON list to override per segment (each item can carry its own prompt, seed and note; unspecified segments fall back to the global prompt).
• base_seed + seed_policy - seeds per segment, with an increment policy by default so segments don't share noise.
• steps, shift_video (12), shift_audio (3), sampler_name (dual_clock_euler), scheduler (native_flow) - the sampling contract, passed through so downstream nodes stay consistent.

The output farm

The orchestrator exposes the whole plan as wire-ready outputs: segment_index, length, context_frames, trim_start_seconds, final_duration_seconds, timeline_start_seconds / timeline_end_seconds, save_context, is_final_segment, prompt, seed, has_next, progress, plan_json, report_json, and the sampling parameters (steps, shift_video, shift_audio, sampler_name, scheduler, sampling_summary). In practice you feed the timing/context outputs into the chain's conditioning and context nodes and let has_next / progress tell you where you are.

Install and honest expectations

Install once for the whole pack: ComfyUI Manager → search MiniMax H3 Audio T8, or git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8 into ComfyUI/custom_nodes/, restart. No extra pip deps, no auto-downloaded weights - bring your own H3 DiT, Qwen3-VL CLIP and video/audio VAEs. Final stitching wants ffmpeg callable from PATH.

This is Experimental and T8-labeled, and the author's philosophy is fail-closed: it would rather stop you than produce a surprise. So when it "blocks" you after a completed manifest, that's a feature - you're done, not broken. And keep your 16GB-card expectations in check: the README is explicit that long chains on one card are the risky regime, and that a chain that runs end-to-end isn't a promise of seamless quality at every seam.