NS Prompt Tuner Load

LLM prompt-rewriters are a blessing and a curse. They turn your rough idea into a strong image prompt - and then they refuse to stop. Every run is a slightly different prompt, so "find the best version" becomes a manual chase where you write down what worked. NS Prompt Tuner Load is the head of a loop that does that chasing for you: it serves the current best system prompt, and each queue run refines it against a design reference, saves the improved version, and the next run picks it up. Turn on Auto-Queue and the loop tunes itself.

It's the more interesting half of a two-node system, so read this alongside NS Prompt Tuner Save. Load is what you put at the top of the graph; Save is the tail that records each improvement.

How it works

Each queue run is one iteration of a tuning loop. Load serves the current best prompt from a state file at output/prompt_tuner/<tuner_id>.json - first run of a new tuner_id seeds v0 from your initial_prompt. You wire that system_prompt into your generator LLM (this pack's NS LLM Chat, or any LLM node), it produces an image prompt, your sampler renders it, and a refiner LLM compares the result against your design reference. Save parses the refiner's CRITIQUE / VERDICT / PROMPT … END PROMPT reply and appends the improved prompt as version v+1. Next run serves that. Auto-Queue = self-improving prompt.

The node's cache fingerprint is the state file, so each saved version actually drives a fresh run instead of serving a cached answer.

Inputs that matter

• tuner_id - one tuning experiment, one state file. Case and punctuation are folded, so Bakery Sheet and bakery-sheet are the same tuner. New experiment = new id.
• guidance - rough notes on what to improve ("palette is good, keep the counts"). Changing it re-opens a converged tuner and steers the next critique.
• initial_prompt - the seed system prompt, used only on the first run of a new tuner.
• max_iterations - lifetime cap on refinements (default 12). Treat this as your spend ceiling on Auto-Queue; 0 means unlimited.
• version_override - -1 = latest (tuning mode). 0 = the original prompt. N = serve vN pinned: the loop never halts, Save records nothing, and repeated runs are fully cached. This is your production/rollback mode.

Outputs: system_prompt (wire to your generator LLM and into the refiner's join), refiner_context (wire into the refiner's prompt), and status (handy into a Preview Any).

Install and setup

Part of the Symbiotica pack. ComfyUI Manager → search "Symbiotica", or:

cd ComfyUI/custom_nodes
git clone https://github.com/symbiotica-ai/comfyui-nodes.git symbiotica
pip install -r symbiotica/requirements.txt

Restart. You'll need LLM API keys (Anthropic, Gemini, etc. - whichever your chat nodes use) via Settings → Symbiotica or env vars. No models to download.

Troubleshooting

• The loop halts with an error - a refiner reply without a PROMPT: section (a refusal or commentary) or without the final END PROMPT line (token-cap truncation) is rejected on purpose. The loop stops rather than saving a poisoned prompt. Rerun.
• It stops after a few runs saying Save is unrecorded - keep the Save node wired and unmuted. It's what records refinements and advances max_iterations. Muted Save = no progress, and the loop tells you so after three unrecorded serves.
• Inconsistent results between runs - fix your sampler seed so the prompt is the only variable. You can't tune a moving target.
• A corrupt state file - the node fails loudly with its path. Repair, delete, or switch tuner_id. State is per-machine: a RunPod install tuning the same id keeps its own lineage.
• It converged too early - change guidance to reopen a converged tuner and push the refiner in a new direction.