ToAPIs GPT Image 2 VIP

GPT Image 2 dropped in April 2026 with two reputations: genuinely good at text rendering and multi-panel consistency, and stubborn about aspect ratios. That second complaint - "how much of a competitor is it, if it still will only do weird aspect ratios?" - is the hole this node plugs.

What it is

This is a cloud-call node. No model loads, no VRAM, no checkpoint. It packages your prompt (plus up to ten optional reference images), fires it at a ToAPIs endpoint running the gpt-image-2-vip mirror of GPT Image 2, and drops the results into your graph as ordinary IMAGE data - text-to-image, or up to ten-way reference image-to-image.

You reach for it for the reason anyone reaches for an API node: the model has no open weights, so local isn't an option at any VRAM. The community's take was "viable Nano Banana competitor" plus "not open source, no interest" - fair, but the text rendering is real, and this gets it into a graph that already has your local upscales.

One honest caveat. This is a reseller route, not OpenAI direct, and the pack - 27 Chinese-authored utility nodes - has essentially no Reddit footprint. The source is readable, but the warning stands: an API node holds your key and phones home by design - exactly the shape of the one custom node that got weaponized in this ecosystem.

How it works

Under the hood it's an HTTP client. It POSTs to /v1/images/generations, gets back a task id - the interface is asynchronous - then polls on poll_interval until the status flips to completed and downloads the result URLs into tensors. Rate limits back off via Retry-After, and transient 5xx responses during polling are tolerated rather than fatal.

Reference images are the interesting part. The endpoint only accepts public http(s) URLs - no base64 - so a local IMAGE gets encoded to PNG and uploaded first, via multipart/form-data to /v1/uploads/images, purely to buy a URL the generator can fetch. PNG goes first to keep alpha; past the API's 10MB cap it silently falls back to JPEG, which flattens transparency. Your references leave the machine twice over.

IS_CHANGED returns the current time, so caching is deliberately off: every queue is a real, billed API call even if you changed nothing.

The inputs that actually matter

• api_key and api_base. Leave the key blank and the node reads TOAPIS_API_KEY from the environment instead. Keep api_base at https://toapis.com; mainland China users switch to https://toapis.cn. A wrong base is the classic 404 here.
• prompt. Up to 32,000 characters. The default is a Chinese placeholder ("a red circle on a white background") - replace it before you queue and wonder why you got a test image.
• size, custom_size, resolution. size covers every preset ratio plus auto; for anything else - 7:4, 1:3 - fill custom_size, which overrides the preset and is capped at a 3:1 ratio. (The 21:9 and 9:21 presets are exempt - vendor presets, not custom ratios.) resolution (1k/2k/4k) decides actual pixels; 4k tops out at 3840 on the long side.
• quality, background, n, and the polling pair timeout / poll_interval. low drafts cheap, medium is the sane default, high is for finals. auto sends no background parameter at all; transparent only pays off if the result stays lossless on the way back. And don't set poll_interval above timeout, or you'll time out before the first poll.
• image_1 (and friends). Connect it and a second port appears, up to ten.

Outputs are images (a list of IMAGE - downstream nodes just run once per image) and response, a JSON string with the task id, resolved settings and result URLs. It's not a terminal node, so wire images into a SaveImage or PreviewImage.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/LaoMaoBoss/ComfyUI-WBLESS.git

Restart ComfyUI, or search ComfyUI-WBLESS in ComfyUI Manager. There's no requirements.txt - the pack vendors its own cozy_comfyui and volcengine libs and falls back to stdlib logging if loguru is missing, so portable builds won't choke. Keep the key in the environment rather than the field:

export TOAPIS_API_KEY=sk-...

That matters when you share workflows - a node field travels with the .json.

When it breaks

The node translates HTTP failures into readable messages. 401 is a bad or missing key, 402 an empty balance, 403 an account that can't touch the model, 404 almost always a wrong api_base for your region, 413 a reference image past 10MB - and if that image had alpha, the JPEG fallback already ate it.

If a reference image seems ignored, check the console: the node warns explicitly when an input was wired but dropped before reaching the backend. If a transparent job comes back opaque, it's the same 10MB JPEG downgrade - shrink the reference first.

And for 4k + high jobs, raise timeout rather than debugging: the error is a timeout, not a bug.