ComfyUI Node
LazyPrompt — Prompt Engineer
A ComfyUI node in vsaan212/LazyPrompt with 25 inputs and 3 outputs.
LazyPrompt — Prompt Engineer
- image
- PROMPT
- PREVIEW
- NEG_PROMPT
◄bypassfalse►
◄user_inputa woman walks through a rain-soaked city street at night►
◄target_model🎬 LTX 2.3 — video, cinematic arc + audio►
◄environmentNone — LLM decides►
◄screenplay_modefalse►
◄creativity0.9 - Balanced Professional►
◄seed-1►
◄invent_dialoguetrue►
◄keep_model_loadedfalse►
◄offline_modefalse►
◄video_length8.00►
◄fps0►
◄env_seed0►
◄modelLM Studio (API)►
◄local_path_8b►
◄local_path_3b►
◄lm_studio_model►
◄lm_studio_ttl5►
◄max_output_tokens900►
◄system_prompt►
◄scene_context►
◄lora_triggers►
◄character►
◄prompt_override_input►
Categoryvsaan212/LazyPrompt
Inputs (25)
| Name | Type | Default | Description |
|---|---|---|---|
| bypass | BOOLEAN | false | When ON, skips the LLM entirely and sends your text straight through. Use for manual prompts or testing. |
| user_input | STRING | a woman walks through a rain-soaked city street at night | Rough idea, sentence, or numbered steps. The LLM expands into a prompt for the selected target model. |
| target_model | COMBO | 🎬 LTX 2.3 — video, cinematic arc + audio | Prompt style: LTX / Wan (video), Flux / SDXL / Pony / SD 1.5 (image). "None" = no JSON template (empty system unless override). Default templates load from lazyprompt/system_prompts.json (restart or R to reload after edits). |
| environment | COMBO | None — LLM decides | Injects location, lighting, and (for video) sound. Random uses env_seed. |
| screenplay_mode | BOOLEAN | false | LTX 2.3 only: screenplay-style character/scene/beats. Ignored for other targets. |
| creativity | COMBO | 0.9 - Balanced Professional | Sampling temperature preset for local HF models and LM Studio. |
| seed | INT | -1-1–2147483647 | LLM seed. -1 = random each run. |
| invent_dialogue | BOOLEAN | true | Video targets: when ON, invent inline dialogue; when OFF, only quoted user dialogue or silence. |
| keep_model_loaded | BOOLEAN | false | Keep the local HF model in VRAM between runs. Off frees VRAM after each run. |
| offline_mode | BOOLEAN | false | ON = no HuggingFace network; local cache / paths only. |
| video_length | FLOAT | 8.000.25–300 | Video duration in seconds (video targets). Frame count for hints = length × fps. Image targets ignore this. |
| fps | INT | 00–60 | 0 = auto from target model (Wan 16, LTX 25, else 24). Set >0 to override. |
| env_seed | INT | 00–2147483647 | Seed when environment is Random. 0 = different pick each run. |
| model | COMBO | LM Studio (API) | Backend LLM: local HF checkpoints or LM Studio REST API (native /api/v1/chat with OpenAI fallback). |
| local_path_8b | STRING | Optional local snapshot path for the 8B model. | |
| local_path_3b | STRING | Optional local snapshot path for the 3B model. | |
| lm_studio_model | STRING | Required when model is LM Studio (API). LM Studio 0.4+ uses /api/v1/chat; older versions fall back to /v1/chat/completions. | |
| lm_studio_ttl | INT | 50–3600 | LM Studio JIT unload: when > 0, unloads the model immediately after each run via /api/v1/models/unload (v1 chat). On OpenAI fallback only, also sends idle TTL in the request body. 0 = leave loaded. |
| max_output_tokens | INT | 90096–16000 | Hard cap on completion length (HF max_new_tokens / LM Studio max_tokens). The pacing hint asks the model for ~one-third of this many tokens by default — raise both together for longer prompts (e.g. 1500 max → ~500 target). |
| system_prompt | STRING | Override system prompt. If empty, uses lazyprompt/system_prompts.json for the selected target. When target is None and this field has text, only that text is used as system prompt and the user message is scene + idea only (no auto augmentation). Default templates load from lazyprompt/system_prompts.json (restart or R to reload after edits). | |
| scene_contextopt | STRING | Authoritative scene/subject text from the vision node or manual paste. | |
| lora_triggersopt | STRING | Injected so the model places these words at the very start of the output. | |
| characteropt | STRING | Subject / character description — always prepended to the LLM user message (including minimal mode and when prompt_override_input is set). Wire from Lazy-subject-and-scene-automation subject_description. | |
| imageopt | IMAGE | Optional. For use with LM Studio (API) and a vision-enabled model loaded in LM Studio. Sends this frame as an image alongside your text (OpenAI-style multimodal chat). Ignored when using local Hugging Face text-only models — use LazyPrompt — Vision Describe + scene_context instead. | |
| prompt_override_inputopt | STRING | When connected/non-empty, replaces user_input for the LLM request (LM Studio API and local HF). Wire from Lazy-subject-and-scene-automation prompt_override. |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| PROMPT | STRING | — |
| PREVIEW | STRING | — |
| NEG_PROMPT | STRING | — |