AtlasCloud Text to Image

What it actually is

GPT Image 2 - OpenAI's 2026 follow-up in the GPT Image line - has no open weights. You can't download it, quantize it, or run it on your GPU. This node is the workaround: it's a thin HTTP client that sends your prompt to atlascloud.ai and pulls the finished image back into ComfyUI as a normal IMAGE tensor. From the canvas it behaves like any other generator - prompt in, image out - but the actual diffusion happened on someone else's server.

The name hides that. AtlasCloud is a pay-per-call inference host in the same shape as fal or Replicate, and this pack is one of its own integrations. So the whole thing to internalize up front: no model files, no VRAM, just a key and a network call. If you were hoping for a local GPT Image, that's not this.

How it works

Grep nodes.py and there's no magic. The node POSTs a JSON payload to /api/v1/model/generateImage with your key in a Bearer header, then polls /api/v1/model/prediction/{id} every couple of seconds until the job reports completed or failed. When it lands, the returned base64 image is decoded back into a torch tensor and handed to ComfyUI like a local sampler's output. It sets enable_base64_output so images come back inline, only sends a seed if it's >= 0, and wraps the whole call in a 5-minute timeout. Failures raise a readable exception instead of silently returning nothing.

The inputs that matter

Required is just three fields: base_url (leave it at https://api.atlascloud.ai unless you're pointing at a self-hosted endpoint - the node normalizes trailing slashes and /api suffixes for you), api_key (the one thing that must be real), and prompt. Everything else is optional, and honestly only two of the knobs get touched in normal use:

• size - an enum of 11 official presets split into 1K (1024x1024, 1024x1536...), 2K (up to 2560x1440) and 3K (3840x2160). It's constrained for a reason: the API only accepts these.
• quality - low / medium / high; higher takes longer and costs more, per the author's own tooltip.

seed (-1 = random) is your reproducibility lever once you hit something you like. output_format is jpeg or png. And extra_params is the escape hatch: paste a JSON object there and it's merged into the request body, which is how you reach API features the node doesn't expose.

Outputs: images (IMAGE, ready to feed a preview or save node) and info (STRING, a small JSON blob with model, size, quality and seed) - handy when you want to log exactly what produced a given frame.

Installing it

ComfyUI Manager → search "ComfyUI-AtlasCloud", or:

cd ComfyUI/custom_nodes
git clone https://github.com/clownvary/ComfyUI-AtlasCloud.git

Then restart ComfyUI. That's the whole install - there's no model download and no dependency hunt beyond torch/pillow/numpy/requests, all of which a stock ComfyUI environment already has. The nodes show up under the AtlasCloud category as "AtlasCloud Text to Image."

Gotchas

• An empty api_key or empty prompt raises immediately, so those errors are easy to read.
• Bad JSON in extra_params throws a "please provide valid JSON" exception.
• It's a closed model, which means OpenAI's moderation applies on their side - the node can't bypass a refusal, whatever a reseller promises.
• Your prompt leaves the machine for a server you don't control. That's the deal with every API node, and this one is no exception.
• Cost is per call and it adds up faster than people expect. A local generation is free after electricity; this is metered every time. Use it because you specifically want GPT Image 2, not as a daily driver.
• One thing the README doesn't shout about: its atlascloud.ai link carries a referral code, so the author earns a cut if you sign up through it. Fine - just know it.