The LLM node calls a language model and outputs text. On a generative canvas
its most common job is prompt expansion: turn “sweet cat” into a vivid,
concrete paragraph an image or video model can do justice to — that’s what its
default system prompt is tuned for. It’s equally happy writing copy,
summarizing, or describing a wired-in image.
|
|
| Inputs |
prompt (text), image (image, optional — for vision tasks) |
| Outputs |
text |
| Param |
Notes |
model |
Claude Opus 5 / Sonnet 5 / Haiku 4.5, GPT-5.1, GPT-4o-mini, Grok 4.5, Gemini 3.1 Flash, Llama 4 Scout, DeepSeek — pick per task. |
prompt |
Fallback prompt when the prompt input isn’t wired. |
system |
The system prompt. Default: expand a short idea into one vivid, generation-ready paragraph. Replace it to change the node’s job entirely. |
prompt_enhancer |
off / generic / cinematic — pre-expansion of the incoming prompt. Usually unnecessary here; the LLM is the enhancer. |
- Prompt amplifier — Text Prompt → LLM → Image/Video Gen. The classic.
Keep your input short and let the model add subject detail, setting,
lighting, mood.
- Style-locked writer — set
system to your brand voice and feed it
briefs; wire the output wherever copy is needed.
- Image describer — wire an image into
image and ask for a detailed
description; feed that to a generation node to riff on an existing look.
- Chain of two — a fast model (Haiku, Flash) to draft, a strong model
(Opus, GPT-5.1) to refine. Each node’s output is inspectable, so you see
what each step added.
