SF VertexAI Imagen 3 Text to Image

Google's Imagen is one of the few frontier image models you can't run locally - there are no weights, and there never will be. If you want its clean, typography-solid output inside ComfyUI instead of on the ImageFX website, SF VertexAI Imagen 3 Text to Image is the bridge. It's a thin wrapper around Vertex AI's Imagen 3.0 API that returns a normal ComfyUI IMAGE tensor, so everything downstream - save, upscale, composite, feed into another model - just works.

That's the real value proposition, by the way: not that the node does anything clever, but that it turns a cloud call into a first-class node in your graph. It's part of comfyui-stillfront, the SF pack, under Stillfront/VertexAI.

How it works

The node uses Google's google-genai library with vertexai=True, which means authentication is your standard Google Cloud Application Default Credentials - the same gcloud auth application-default login you'd use for any GCP project. You give it a project_id and a location (default us-central1), it calls models.generate_images, and it converts the returned bytes back into a tensor. Simple, but you need a Google Cloud project with the Vertex AI API enabled and a billing account. No GPU, no model downloads, no local VRAM.

The inputs that matter

Most of the fields are what you'd expect from an image generator, with two model-specific catches:

• model - three variants: imagen-3.0-generate-002 (default, best quality), imagen-3.0-generate-001, and imagen-3.0-fast-generate-001 (quick and dirty). The catches: negative_prompt only works on the 001 models, and enhance_prompt only works on the 002 model. The author wired both in but they're mutually exclusive by Google's own API design - pick your model, then the corresponding feature.
• aspect_ratio - five options (1:1, 16:9, 9:16, 4:3, 3:4), mapped to Imagen's fixed resolutions.
• num_images - 1 to 4, returned as a batch tensor.
• seed - for reproducible results; it has a "control after generate" widget, so it auto-randomizes after each run unless you lock it.
• Optional: safety_filter_level and person_generation - Google's policy dials. Defaults are BLOCK_MEDIUM_AND_ABOVE and ALLOW_ADULT, which is the middle of the road.

Output is a single images output (IMAGE). Wire it to a preview or save node.

Installing it

Same pack install as everything else in this collection:

cd ComfyUI/custom_nodes
git clone https://github.com/razvanmatei-sf/comfyui-stillfront
cd comfyui-stillfront
pip install -r requirements.txt

Then restart. The requirements.txt pulls in google-genai and google-cloud-storage, which are the heavy bits for the Vertex nodes. Before you run, make sure your GCP project has the Vertex AI API enabled and you've authenticated:

gcloud auth application-default login

You can also set GOOGLE_CLOUD_PROJECT (and GOOGLE_CLOUD_LOCATION) as env vars so you don't have to type the project ID into every node.

Troubleshooting

• "No valid images were returned by the API." This is the big one, and it's almost always the safety filter. The node literally prints the rai_filtered_reason to the console when an image is blocked. Loosen safety_filter_level to BLOCK_NONE if your project's policy allows, and restate the prompt - this is Google's censorship, not a bug.
• "Imagen API call failed." Usually auth. Check that ADC is set up (gcloud auth application-default login) and that your project ID is right - the node needs it if you haven't set GOOGLE_CLOUD_PROJECT.
• Negative prompt silently ignored. You're on the 002 model. Switch to a 001 variant.

One honest caveat: Imagen images carry Google's invisible SynthID watermark, and the service is aggressively moderated compared to local generation. If you're building an uncensored pipeline, this is the wrong tool - it's for when you want Imagen's specific look and are fine with Google's rules.