Prompt Enhance (MRLN)

Rewrite a prompt with an LLM under a target-model system prompt. ONE wire does it: the Prompt Template node's llm output carries the rendered prompt plus the selected profile's system prompt, which tells the LLM how the TARGET image model wants its prompts (prose for KREA/FLUX, tags for SDXL/Pony, ...). The optional prompt input enhances any other STRING instead and wins when both are wired. Backends: local Ollama / LM Studio (URLs in the Composer's Settings tab) or cloud Anthropic / OpenAI / Gemini / OpenRouter (API keys stored there, server-side only — never in widgets). Deterministic per seed where the backend supports it, cached per input so re-queues never re-call, and a failing backend passes the original prompt through instead of killing the render (switchable). Ollama frees its VRAM right after the call by default so the sampler gets the GPU back. LoRA trigger words riding the llm wire are PROTECTED: they are demanded verbatim in the system prompt and verified after the rewrite — any span the LLM dropped or mutated is re-injected, so an enhancement can never disarm a LoRA.

By master-merlin·Created 16 days ago·Updated 7 days ago· 2

prompt
report

◄backend▾►

◄model►

◄temperature0.20►

◄seed0►

◄max_tokens512►

◄timeout60►

◄free_vram▾►

◄on_error▾►

◄llm—►

◄prompt—►

◄system►

CategoryMRLN/prompt

Inputs (11)

Name	Type	Default	Description
backend	COMBO		LLM backend. Local: Ollama / LM Studio (URLs in the Composer's Settings tab). Cloud backends need an API key stored there — in the browser only keyed ones are listed.
model	STRING		Model name, e.g. 'gemma3:12b' (Ollama) or an LM Studio model id. Required for Ollama; LM Studio falls back to its loaded model; cloud backends fall back to a sensible default. In the browser this becomes a dropdown of installed models plus pull suggestions Ollama downloads on pick.
temperature	FLOAT	0.200–2	Sampling temperature — keep low for faithful rewrites.
seed	INT	00–18446744073709550000	LLM seed for reproducible rewrites (where the backend supports it). 0 derives a stable seed from the prompt + system text, so identical inputs enhance identically.
max_tokens	INT	51216–8192	Generation cap for the rewrite. Auto-raised when the input is longer than the cap allows — a keep-everything rewrite can never be shorter than its input (the report notes when this happens).
timeout	INT	605–600	Seconds to wait for the backend before giving up.
free_vram	COMBO		Ollama keep_alive: 'after call' unloads the LLM immediately so the diffusion model gets the VRAM back (recommended on one GPU); 'keep 5m' keeps it warm for rapid iteration; 'always keep' pins it loaded until Ollama stops (second GPU / big VRAM). LM Studio manages its own lifetime.
on_error	COMBO		When the backend is unreachable or errors: pass the ORIGINAL prompt through (render never dies, report says why) or raise and stop the queue.
llmopt	STRING		The Prompt Template node's llm output — the single wire: {target, prompt, protect, system, params}. It carries the rendered prompt, the profile's system prompt, and the LoRA trigger words that are enforced verbatim (dropped ones are re-injected).
promptopt	STRING		Optional override: enhance this STRING instead of the prompt carried inside the llm input (wins when both are wired). Use it to enhance text from any other node.
systemopt	STRING		System prompt override. Empty = use the llm input's system prompt; set both and this one wins (template guides, user decides).

Outputs (2)

Name	Type	Description
prompt	STRING	The enhanced prompt (or the original on pass-through).
report	STRING	What happened: backend, model, seed, cache/VRAM state, or the pass-through reason.