Griptape Create: Image from Text

Griptape Create: Image from Text is where this pack stops being an LLM wrapper and becomes an image generator. Give it a prompt and it produces an image through Griptape's image-generation tasks - but the interesting part is how it plugs into everything else. It's the node in the README's signature example: three image descriptions get merged into one string, and that merged inspiration becomes the prompt for this node. It's not a standalone "type a prompt, get a PNG" tool; it's the image output stage of an agentic pipeline.

What it is, honestly

ComfyUI already does image generation far better than this node will - if you're after diffusion quality, use a proper sampler and checkpoint. What this node gives you instead is LLM integration. The prompt can come from an agent, from a merged pile of image descriptions, or from any text node in your workflow, and the generation happens through a Griptape driver rather than the local diffusion stack. That's the point: closing the loop between "an LLM reasoned about what to draw" and "an image exists."

How it works

It runs a PromptImageGenerationTask inside a Griptape pipeline. Which provider draws the image depends on the driver you attach:

• No driver attached → it falls back to OpenAI's dall-e-3 (quality hd, style natural), which needs an OPENAI_API_KEY from Settings → Griptape or your environment.
• A driver attached → whatever the driver is: OpenAI, Amazon Bedrock Stable Diffusion, Bedrock Titan, Leonardo.AI (and the pack also wires in Black Forest / Flux drivers). The README's "Generate Images" section lists them.

The generated image is written to ComfyUI's temp directory, loaded back as a normal IMAGE tensor, and handed to you with its file path.

Inputs

• STRING (required, multiline) - the prompt. This is what gets generated.
• input_string (STRING) - appended to STRING with a newline; use it for dynamic text from other nodes.
• agent (AGENT) - an optional agent; if the agent's config carries an image-generation driver, that's used.
• key_value_replacement (DICT) - {{ key }} placeholder substitution in the prompt.
• driver (DRIVER) - the image-generation driver node. Attach one to pick your provider and model.

Outputs

• IMAGE - the generated image as a ComfyUI tensor, ready for display/save nodes.
• AGENT - the agent that ran the task, for chaining.
• file_path (STRING) - where the image was written (temp dir).

Installation

Part of ComfyUI Griptape Nodes:

• ComfyUI Manager: search "Griptape" → install.
• Manual: cd ComfyUI/custom_nodes && git clone https://github.com/griptape-ai/ComfyUI-Griptape, restart ComfyUI.

You need a working image-generation driver or an OpenAI key, plus the pack's shared install (griptape[all], python-dotenv, openai). The usual torch-conflict troubleshooting in the README applies if the pack install shakes up your environment.

Gotchas

• Cost. Each generation is a paid API call (dall-e-3 is not free). If you're used to local generation, the billing surprise is real.
• Don't expect local-diffusion quality control - no seed control, no sampler, no steps. It's a driver-based API call, tuned by your prompt, not by denoising knobs.
• The output lands in the temp directory; if you want it kept, save it from the IMAGE output rather than relying on file_path to survive restarts.

Use it when your workflow needs an image that's driven by LLM reasoning. For everything else, ComfyUI's native samplers are the right tool - and this node is the one that bridges the two worlds.