LazyPrompt — Prompt Engineer

Type a rough idea into LazyPrompt - Prompt Engineer and it runs a local LLM that turns it into a full prompt shaped for whatever model you're actually sampling. "A woman walks through a rain-soaked city street at night" becomes a proper cinematic block for LTX, or a tag-dense paragraph for Pony, depending on the target_model you pick. If you've ever stared at a blank prompt box before a Wan render, this is the node that fixes that.

Why this exists: in 2026 prompting is model-specific in a way it wasn't in the SD 1.5 era. A booru tag list that sings on Pony is garbage on Z-Image, and an instruction-style sentence that works on LTX gets mangled into Flux. The whole trick of LazyPrompt is that it doesn't just expand text - it swaps in a different system prompt template per target model, so the same rough idea gets rewritten in the right dialect for LTX 2.3, Wan 2.2, Flux, SDXL, Pony, or SD 1.5. That per-model templating is genuinely the hard part, and it's what most ad-hoc "prompt enhancer" nodes get wrong.

How it works

Three backends, all local:

• LM Studio (API) - the default. This is the one people trip on: the name sounds like a hosted service, but there's no key and no cloud. It talks to a model you run yourself in LM Studio over its native /api/v1/chat endpoint, with an OpenAI-compatible fallback for older versions. Set lm_studio_model to the exact model id as shown in LM Studio - that field is required and it's the #1 "nothing comes back" cause.
• Local Hugging Face - two small abliterated models, NeuralDaredevil 8B and Llama-3.2 3B, downloaded automatically on first run. Fine for text, but they're text-only: the image input is ignored on this backend.
• Bypass - flip the bypass toggle and the LLM is skipped entirely; your text passes straight through. Great for testing whether the LLM is the problem or the fix.

Beyond the system template, you get environment - 73 presets like Beach - golden hour or Night sky - open field that inject location, lighting, and (for video) sound into the expansion. creativity is the temperature: 0.7 stays literal, 1.1 gets artistic. For video targets, video_length and fps feed pacing hints so the LLM writes to a real duration instead of guessing.

The inputs and outputs that matter

The three you'll actually touch on day one: user_input (your rough idea), target_model (which template/pacing rules), and model + lm_studio_model (the backend). After that: seed (-1 = random), max_output_tokens (default 900; the pacing hint asks for roughly a third of it, so raise both together for long cinematic prompts), and keep_model_loaded - leave it off unless you hate the reload time, because a resident 8B LLM eats VRAM you may want back.

Worth wiring from elsewhere in this pack: character (subject description, so your character file always reaches the LLM), scene_context (from Vision Describe), and prompt_override_input - when connected it replaces user_input entirely, which is how the Lazy-subject-and-scene-automation node's prompt_override takes over the LLM's marching orders.

Outputs are three strings: PROMPT (the expanded positive - wire this to CLIP Text Encode), PREVIEW, and NEG_PROMPT. There's no model/tensor output; this node is purely the text half of your graph.

Install

Install the whole pack once: in ComfyUI Manager, search vsaan212/Vsaan212-workflow-utilities and install, or

cd ComfyUI/custom_nodes
git clone https://github.com/vsaan212/Vsaan212-workflow-utilities

then restart. The heavy lifting is in the pack's requirements.txt - transformers, accelerate, torch, qwen-vl-utils - and the first run against a local HF model downloads several GB of weights. With LM Studio you skip that entirely.

Where people get burned

• lm_studio_model empty or wrong id → silent nothing. LM Studio has to be running with its server up (default port 1234).
• Output is your input, unchanged → bypass is on, or prompt_override_input is wired and empty-ish.
• The image input did nothing → you're on the local HF text backend. Feed frames through LazyPrompt - Vision Describe → scene_context instead; image only works with a vision model loaded in LM Studio.
• VRAM won't come back → you left keep_model_loaded on. Turn it off or drop the LazyPrompt - Unload local model node into the graph.

System prompt templates live in lazyprompt/system_prompts.json and you can edit them; press R (or restart) to reload after changes.