OpenAI Image Generation

OpenAI Image Generation is the image side of the ERPK OpenAI section: prompt in, IMAGE tensor out, no browser tab needed. Default model is gpt-image-2, OpenAI's current flagship with 4K output and strong multilingual text rendering - which is the whole reason to pick this over a local checkpoint when you need legible text in the frame, like posters, UI mockups, or logos.

Two things make this node genuinely more useful than a bare API call. First, it outputs two values: image (the IMAGE tensor) and revised_prompt (the STRING of what OpenAI actually generated, after its internal prompt rewriting). That second output is a gift for debugging - when the result isn't what you asked for, the revised prompt shows you exactly how the model reworded your request. Wire it to a text preview and watch what's happening.

Second, the background input supports transparent. That's a real workflow unlock: generate a logo or product render directly on a transparent background, no chroma-key or background-removal step. It requires an output format that supports transparency, so pair it with png/webp handling downstream.

The inputs that matter

• prompt - what you want
• size - default 1024x1024, but gpt-image-2 takes arbitrary sizes: both edges divisible by 16, aspect ratio between 1:3 and 3:1, max edge 3840px. 1536x1024 is the standard landscape option
• quality - auto/low/medium/high for the GPT Image family
• n - 1 to 10 images per call, stacked into a batch
• background - auto or transparent (the sleeper feature)
• moderation - auto (default safety filters) or low (more permissive content)

Plus the shared plumbing: seed for best-effort reproducibility, client for the optional OpenAI API Config node.

Install

Part of the ERPK Collection:

cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk && pip install -r requirements.txt

Requires openai>=2.32.0 and an OpenAI key in ERPK Settings. Or "ERPK Custom Nodes" via ComfyUI Manager, restart, done.

Troubleshooting

The one real trap here is DALL-E. The README is blunt: DALL-E 3 shuts down on 2026-05-12, and the hd/standard quality values are legacy DALL-E options kept only for compatibility. If you load an old workflow with a DALL-E model selected, expect it to break or behave oddly - switch to a GPT-Image model. transparent background on a model or format that doesn't support it will also fail or come back opaque; the tooltip calls that out explicitly. And as with any hosted image model, n > 1 returns a batch, so use an Image node's batch controls to flip through your four or ten candidates.