GPT-Image 1.5 (fal)

This is OpenAI's GPT-Image 1.5 - the December 2025 refresh of the ChatGPT image model - dropped into ComfyUI as a node. It's closed, it's API-only, and it does not run on your machine. GPTImage15 is a thin wrapper (from gokayfem's ComfyUI-fal-API pack, the same person behind VLM Nodes) that hands your prompt to fal, which calls OpenAI, and returns the image as a native ComfyUI IMAGE.

Why bother when you have Flux or Z-Image locally? Two reasons. GPT-Image is genuinely good at text rendering and at following a literal instruction, and it does something most local models make you fight for: clean transparent backgrounds. If you make stickers, logos, product cutouts, or UI assets, this node earns its keep.

How it works

GPT-Image is natively multimodal - it reasons about the prompt as language before it paints. You're not sampling a diffusion model here; you're calling a hosted endpoint and paying per image. Everything happens on OpenAI's servers via fal, so there's no VRAM cost and no checkpoint to manage - just latency and a bill.

The inputs and outputs that matter

The three settings you'll actually touch:

• prompt - write it like an instruction. GPT-Image handles "a red enamel pin of a fox, centered, on a plain background" better than keyword salad.
• background - auto / transparent / opaque. Set it to transparent and you get a real cutout with no matte-removal step. This is the feature.
• quality - low / medium / high. This is also your cost dial: higher quality costs more per image, so drop to medium while you're iterating.

The rest: image_size gives you three shapes (1024x1024, 1536x1024, 1024x1536); num_images runs 1–4; output_format is png / jpeg / webp (use png if you want that transparency to survive); and sync_mode controls whether the call blocks. Output is a single IMAGE.

How to install it

Via ComfyUI Manager: search ComfyUI-fal-API, install, restart. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI-fal-API.git
pip install -r ComfyUI-fal-API/requirements.txt

No models download - it's an API pack. Then get a fal key from fal.ai and either paste it into config.ini (copy config.ini.example first) under [API], or export FAL_KEY=your_key. Restart, and find the node under FAL.

Common issues & troubleshooting

• No key, no output. Set FAL_KEY and restart. Auth failures are the number-one first-run problem.
• A prompt gets refused. OpenAI's content policy runs server-side, and the pack shows you fal's actual error text - if it's a policy rejection, it'll say so. Rephrase or move on.
• Transparency isn't sticking. Set background to transparent and output_format to png. A JPEG can't hold an alpha channel.
• It's slower/pricier than you expected. It's a real OpenAI call billed through fal. The pack's result cache serves an identical call for free on repeat (bypass with force_rerun), and the spend guard will stop you before you blow past a budget - worth enabling.
• Windows Portable missing fal-client. python_embeded\python.exe -m pip install fal-client.

It's not the node for high-volume, unlimited generation - that's what your local models are for. It's the one you keep in the toolbox for the specific jobs GPT-Image is unreasonably good at: legible text and clean transparency.