MiniMax H3 Background Start / 后台长视频启动 (EXP/T8)

Long video runs are long, and the worst failure mode isn't the OOM - it's the OOM that takes down the whole chain and leaves you rebuilding segment 1 of 8 from scratch. This node is the first half of the fix: it registers the current prompt with the background system before model execution starts, so when an upstream error or OOM happens, the run can be retried from the known state instead of losing everything. The pack's own regression notes show a real 256×256×22 H3 run being cancelled at 1/4 steps and returning VRAM to baseline+90MiB - the recovery path is not hypothetical.

It's in T8/MiniMax H3/Long Video/Experimental and it's the entry point of the explicit background workflow. "Explicit" is the key word: nothing auto-accepts unless you opt in.

How it works

You set chain_id (default my_h3_long_video), and the node registers the current prompt under it before any model work runs. execution_mode is the safety dial: review_only (default) is non-mutating - it registers and observes. auto_accept_and_continue accepts every successfully rendered candidate without human review and queues the next segment. That last one is a big switch and the tooltip says it plainly: it must be selected explicitly. The default world is "look at every segment," which is where you want to live until you trust the chain.

max_retries (default 1) is extra attempts per failed segment, and retry_delay_seconds (default 2) spaces them out. Notably, retries never reduce resolution - the tooltip is explicit that resolution is never silently dropped to make a retry fit. release_policy (default clear_execution_cache) is applied after every accepted segment, including pause and final boundaries: clear_execution_cache resets ComfyUI execution tensors and soft cache; unload_all_models is stronger but globally unloads every ComfyUI model, not just H3.

The outputs - chain_id, auto_accept, job_id, background_state_json - connect to the background workflow (the controller and the terminal Auto Accept & Continue node), per the description.

The inputs that matter

• execution_mode - review_only vs auto_accept_and_continue. The single most important decision in the long-video family.
• chain_id - the identity that ties all segments and the manifest together.
• max_retries / retry_delay_seconds - your recovery knobs. Keep them modest.
• release_policy - clear_execution_cache is the balanced default; unload_all_models is the heavy hammer.

Outputs

chain_id and job_id (strings, for the controller), auto_accept (BOOLEAN mirroring the mode), and background_state_json (the state the terminal node reads).

Installing it

Pack install: ComfyUI Manager → "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.

Common issues

It accepts segments you never reviewed. You selected auto_accept_and_continue. That's the mode's contract, not a bug. Switch back to review_only if you want the human gate back - it's non-mutating by design.

Nothing seems to retry after an OOM. Check that max_retries isn't 0 and that the prompt was actually registered (the background_state_json output). Registration happens before execution - if the node sits downstream of the sampler, it's too late to help.