MiniMax H3 Creator × Long Video Run Select / 按后台进度选择镜头 (Advanced/T8)

Once your Creator workspace is bound to a Long Video background chain (that's the Background Start node's job), something has to decide which shot and which seed variant to render next, based on what actually completed in the background. That's MiniMaxH3CreatorBackgroundRunSelectT8Advanced. It reads the bound chain's accepted_count and retry_count, maps them onto your Creator plan, and emits the next shot's prompt packet, length and deterministic seed. It reads progress; it never drives the queue.

The description draws the boundary hard: "It never owns the queue; paused, failed, cancelled, detached or review-only states block downstream generation." So this node is the decision, not the execution - the execution stays in the Long Video background controller where it belongs. If the background state says the chain is paused or the binding is detached, ready comes back false and downstream generation is blocked instead of blindly guessing.

How it decides

• workspace (custom H3_T8_CREATOR_WORKSPACE) - your shot plan.
• background_state_json - the live state from the Background Start node's output. This is the only source of truth it trusts.
• variant_policy (retry_as_variant_clamped) - how retries map onto variants. The default treats a retry as another variant, clamped to the workspace's bounds, so seed selection stays deterministic and reproducible.

What comes out

The useful set: prompt_packet (custom type, ready for conditioning), compiled_prompt, negative_prompt, length, seed, run_position, variant_index, attempt_number, ready (BOOLEAN gate), action (STRING - what the plan says to do next: render, review, retry…), shot_json and report_json. Wire ready into a gate and seed into your sampler and you have a loop that walks the entire shot list based purely on what the background queue confirms it finished - including retries on rejected shots, which is the "long video" dream that usually turns into a hand-cranked nightmare.

The honest caveats

This is experimental and deliberately conservative. It picks shots and seeds; it can't tell you whether a completed candidate is good - that's the human review step (or the pack's synchronized A/B preview nodes). And because it's bound to the Long Video background schema, the state string has to be a real one from the matching chain; pasting a stale or fabricated state gets you ready=false, not a weird render. The README's validation only claims mechanical correctness on a real 256×256×22 chain, so expect to validate the loop on a small canvas first.

Install

Standard pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8

Restart ComfyUI. No extra dependencies. It lives in T8/MiniMax H3/Studio/Experimental, wired between Background Start and your conditioning/sampler in the 2026-08-23_H3_Creator_Long_Video_Background_Bridge_Advanced_EXP.json workflow - copy that graph, swap in your own shots, and let the queue do the walking.