MiniMax H3 Creator Workspace Shot Select / 工作区镜头选择 (Advanced/T8)
Pick the shot and seed you're about to burn VRAM on
- workspace
- prompt_packet
- compiled_prompt
- negative_prompt
- length
- seed
- shot_json
- report_json
The Creator Workspace is a plan; this node is the moment the plan becomes a render. MiniMaxH3CreatorWorkspaceShotSelectT8Advanced takes a compiled workspace plus a run_position and a variant_index, and hands you back a ready-to-use prompt packet, the compiled prompt text, a length, and - the part that matters when you're about to sit through an H3 sample - the exact deterministic seed for that variant. No hidden queue execution, no sneaky extra jobs. You pick, it tells you what you picked, and then you wire the output into your existing H3 Conditioning/Sampler nodes.
Why you'd reach for it
This is the "what am I about to render, exactly?" node. In a variant-heavy workflow you've got, say, 3 shots × 4 seeds and you only want to burn GPU on seed #2 of shot #1 this pass. Instead of rebuilding prompts or hand-typing seeds into samplers, you flip run_position and variant_index here and everything downstream - the compiled_prompt, the length, the seed - updates as one coherent unit. It's the deterministic bridge between the plan and the sampler, which is exactly the kind of plumbing the KB's node-plumbing notes tell you to expect in production graphs: one authoritative source, fanned out to many consumers.
Inputs and outputs that matter
Only three inputs, and two are plain integers: workspace (the H3_T8_CREATOR_WORKSPACE from Creator Workspace), run_position (which shot in the run window, 0-based), and variant_index (which seed variant, 0–63).
The useful outputs:
prompt_packet(H3_T8_PROMPT_PACKET) - the compiled conditioning packet; feed it to the pack's Conditioning node.compiled_prompt/negative_prompt- text versions, great for a Preview or a Note before committing.length- frame count for this shot; keep it aligned with H3's grid values.seed- the deterministic seed for this variant. Wire it to your sampler and you get the same frame every time for the samevariant_index.shot_json/report_json- the shot's full record and the audit trail.
Installing it
Part of the MiniMax H3 Audio T8 pack (T8mars/T8star's H3 kitchen sink, GPL-3.0). Via Manager search MiniMax H3 Audio T8, or:
cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8
Restart after. No mandatory extra pip packages, and no models are auto-downloaded - the H3 weights, CLIP, VAE and LoRAs go in models/ yourself. Keep in mind the H3 weights' Community License excludes the US, EU, UK and South Korea.
Where people get tripped up
Two things. First, this node is an output node - it ends a branch, so don't expect it to also run the sampler. You still wire prompt_packet/seed/length into the actual generation nodes. Second, variant_index and run_position are both 0-based and independent; mixing them up is the classic "why is seed 0 rendering again" moment. And if seed doesn't change when you bump variant_index, check that the shot's Shot Override actually has variant_count > 1 - a one-variant shot only has one seed to give you.
Inputs (3)
| Name | Type | Default | Description |
|---|---|---|---|
| workspace | H3_T8_CREATOR_WORKSPACE | — | |
| run_position | INT | 00–100000 | — |
| variant_index | INT | 00–63 | — |
Outputs (7)
| Name | Type | Description |
|---|---|---|
| prompt_packet | H3_T8_PROMPT_PACKET | — |
| compiled_prompt | STRING | — |
| negative_prompt | STRING | — |
| length | INT | — |
| seed | INT | — |
| shot_json | STRING | — |
| report_json | STRING | — |