Prompt Enhance (MRLN)

LLM prompt-enhancers are everywhere now, and the community has learned their failure modes the hard way: they add detail you never asked for, they emit chat instead of a clean prompt, and they eat your trigger words. Prompt Enhance (MRLN) is built around those specific failures rather than pretending they don't exist. It's the optional LLM rewrite stage of this pack, and it takes exactly ONE wire - the llm output of the Prompt Template node.

That wire is the interesting part. It doesn't just carry the rendered prompt; it carries 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 - plus the LoRA trigger words riding the prompt. Those trigger words are protected: demanded verbatim in the system prompt, verified after the rewrite, and any span the model dropped or mutated is re-injected. An enhancement can never disarm a LoRA, which is more than most enhancers on the market can claim.

How it works

Backends are local Ollama / LM Studio (URLs configured in the Composer's Settings tab) or cloud Anthropic / OpenAI / Gemini / OpenRouter (API keys stored server-side, never in node widgets - in the browser, only keyed backends are even listed). The rewrite is 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 with on_error.

Ollama gets special treatment: free_vram: after call (the default) unloads the LLM immediately so the diffusion model gets the GPU back. On a single card that's the difference between an enhancer you can live with and one that OOMs every render.

The inputs that matter

• backend and model - model is required for Ollama; LM Studio falls back to its loaded model; cloud backends fall back to a sensible default.
• temperature - keep it low. 0.2 is the default, and it's the right number for faithful rewrites.
• free_vram - after call is the one-GPU recommendation. keep 5m if you're iterating hard, always keep only on a second GPU.
• on_error - pass through means a dead backend never kills a render; raise stops the queue loudly instead.

There are two optional inputs worth knowing: prompt enhances any other STRING and wins when both it and llm are wired (use it on text from any node), and system overrides the profile's system prompt when you want to steer the rewrite yourself.

Outputs: prompt - the enhanced prompt, or the original on pass-through - and report, which tells you backend, model, seed, cache/VRAM state, or the pass-through reason.

The honest take

The README says it plainly: this node is best on thin hand-typed prompts and tag→prose conversion - the curated library usually renders better un-rewritten. So it's not a "make everything better" button. It's for when you have a rough idea you want expanded for a specific target model, or a tag list you want turned into prose for a FLUX/KREA-class model. The max_tokens handling is a nice touch: a keep-everything rewrite can never be shorter than its input, so the cap auto-raises rather than silently truncating the tail of a long prompt (the report notes when that happens).

How to install it

Same pack, same note: ComfyUI Manager can't find this pack yet - the registry flagged v0.1.1 on six severity: info matches (including the LLM backend calls, which is part of why it's flagged), so it's absent from Manager's search. Review hiccup, not a broken node. Manual install:

cd ComfyUI/custom_nodes
git clone https://github.com/master-merlin/comfyui-mrln-nodes ComfyUI-MRLN-Nodes

Restart ComfyUI. No extra Python dependencies - requirements.txt is deliberately empty. You'll want Ollama running if you use the local path, but that's a separate install.

Common issues

• The render works but the prompt is unchanged - backend unreachable with on_error: pass through. Check the report output; it names the reason.
• VRAM contention on one GPU - Ollama and the diffusion model fighting over the card. Set free_vram to after call and don't keep the LLM resident.
• "Model name required" - Ollama needs a real model id like gemma3:12b; it won't guess.

A dead backend never breaks a render by default, which is the right default - but it also means "it silently did nothing" is the failure mode to watch for.