Griptape Driver: OpenAI Image Generation

If you want your Griptape agent to not just talk about an image but produce one, this is the node. The OpenAI Image Generation driver wires an image-generation backend into your agent, so a Griptape Create: Image from Text task can call out to OpenAI's image models (dall-e-2, dall-e-3, or gpt-image-1) instead of only reasoning about text.

Set your expectations up front: this is not local generation. It doesn't touch your checkpoint or your GPU. It's a paid API call that returns a finished image, which Griptape hands back to the workflow. That's a feature in agent-land - your local diffusion setup stays free for what it's good at, and the agent gets a cheap way to illustrate its output - but it's also the thing people miss when they grab this node expecting a ComfyUI-quality render. The output is good, but it's dall-e, not your fine-tuned SDXL.

How it works

The node wraps Griptape's OpenAiImageGenerationDriver. You pick a model and a size, it reads OPENAI_API_KEY from the environment variable you name, and it hands you a DRIVER object. That driver then plugs into an agent config's image-generation slot - Griptape Agent Config: Custom Structure accepts it, or you can use it in the OpenAI-flavored configs.

One genuinely thoughtful detail hides in the source: the node auto-corrects invalid size/model combos instead of erroring. Pick dall-e-2 and a portrait size like 1024x1792? It silently downgrades you to 1024x1024 because dall-e-2 doesn't do non-square. Pick dall-e-3 with 256x256 or 512x512? Back to 1024x1024, because dall-e-3's smallest size is 1024. Handy, but it means the size you selected may not be the size you get - check the result if dimensions matter.

The inputs that matter

• image_generation_model - dall-e-2, dall-e-3 (default), or gpt-image-1. gpt-image-1 is the newest and best, but it's also the most expensive; dall-e-3 is the sane default for agent output.
• size - 256x256, 512x512, 1024x1024, 1024x1792, 1792x1024. Remember the auto-correction above: not every size works with every model.
• api_key_env_var - the name of the env var holding your key, default OPENAI_API_KEY. Never paste the key itself here.

Output is a single DRIVER of type DRIVER, which feeds the config. From there, a task node like Griptape Create: Image from Text uses it to actually render.

Install and keys

Same pack as the rest - ComfyUI Manager (search "Griptape") or:

cd ComfyUI/custom_nodes
git clone https://github.com/griptape-ai/ComfyUI-Griptape

Set OPENAI_API_KEY under Settings → Griptape in ComfyUI (or in your environment). Note the pack's requirements.txt pins an openai version and pulls in griptape[all], which can fight ComfyUI's torch install on Nvidia - reinstall torch from the cu121 index if you see import or runtime errors.

Common gotchas

The usual suspects apply: enter the env var name in api_key_env_var, not the secret. And because this is an API call, it bills you per image and can silently fail on rate limits - the agent will just report an error for that task. If your workflow is a one-off illustration, dall-e-3 at 1024 is the cost/quality sweet spot; save gpt-image-1 for when you actually need its quality jump.