T8 Creative Candidate Lab（多方案实验室）

Most prompt "enhancers" give you one answer. T8 Creative Candidate Lab（多方案实验室） is the node that goes the other way: one LLM request, and it comes back with 2–4 structurally different directions - not four rewordings of the same idea, but four different creative mechanisms. Different premise beats, different camera grammar, different payoff. It's the "explore the space before you commit" node, and it's the recommended start of the suite's iteration flow.

The trick that makes it cheap is the request design. The system prompt demands genuinely divergent creative axes and a compact JSON shape - name, creative axis, prompt, two strengths, two risks, plus a comparison line - and explicitly forbids the model from scoring itself. The scoring happens locally, after the call, with deterministic text heuristics (the source runs the candidate text past regexes for causality, camera control, continuity, media references, temporal feasibility, production risk). That's a deliberate honesty move: the soft scores are advisory, never a quality verdict, and they never block a non-empty candidate. The author's own acceptance notes call this "deterministic advisory scores, not objective art quality."

The inputs

• concept (required) - your creative request.
• creative_brief - optional, from the Director, so all candidates stay inside your locked boundaries.
• candidate_count - 2, 3, or 4 (default 3). All from the same one request.
• divergence - 低 / 中 / 高: same-strategy micro-variations, different creative mechanisms, or clearly different directorial routes. The middle default is the sweet spot.
• must_keep - facts the model can't touch.
• model_target, rewrite_mode (defaults to creative here, which suits the job), output_language, seed, plus api_key / provider_config.

What comes out

• candidates_json - the full payload, including each candidate's soft scores and risks, and the scoring_notice explaining that the scores are local heuristics.
• comparison_report_json - the comparison lines and scoring notice, separated out.
• candidate_1 through candidate_4 - the prompts as plain STRINGs. Unused slots come back empty.

Wire candidates_json into T8 Candidate Selector to choose, or skip the selector and grab candidate_2 directly. For a bigger pipeline, the selected prompt feeds T8 Storyboard Pack's concept.

Installing

Part of the comfyui-minimax-h3-prompt-enhancer-T8 pack (T8 = T8mars, a Bilibili/YouTube ComfyUI creator - the pack is Chinese-ecosystem, so expect the README to be your best friend). Manager: search "MiniMax H3 / Seedance 2.0 / Music 3 Prompt Enhancer (T8)", or:

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

Restart and Ctrl+F5. The single call rides the pack's shared providers - Seedance NZ by default, or a T8 LLM Provider Config wired into provider_config for the AI Workshop, an OpenAI-compatible endpoint, or a key-free local Qwen GGUF. Cloud mode adds no dependencies.

Gotchas

If the model returns non-JSON text, the node doesn't panic - it wraps the raw text as a single "上游原始方案" candidate so you never lose a non-empty result. And don't expect the soft scores to pick your favorite for you; they're tie-breakers and eyeball-checks, not rankings from heaven. Since one call buys you the whole menu, the habit is: set divergence high, read the comparison, pick with the Selector, and only then spend on the real enhancer.