HOTCUT · GPT Image 2

If you've ever fought Flux or SDXL for an hour over a neon sign with actual letters on it, you know the pain this node exists to remove. HotcutGptImage2 sends your prompt to a cloud image model that HOTCUT exposes as GPT Image 2 - the OpenAI family that's genuinely good at text, Cyrillic and faces - and drops the result back into your graph as a normal IMAGE. It's the flagship node of the comfyui-hotcut pack, and the reason the pack exists.

Nothing runs locally. You type a prompt, pick a resolution, and the node does an HTTP dance: submit a task to POST /api/v1/images (model: gpt-image-2, mode: text-to-image), poll GET /api/v1/images/:task_id every couple of seconds, download the result, decode the bytes into an IMAGE tensor. It's billed in HOTCUT's "flames" (огни), which you buy in rubles at hotcut.ru - no foreign card, no subscription, which is the whole point of the service. Your GPU never wakes up, and the pack's dependencies are just requests, Pillow and numpy.

Three inputs actually matter:

• prompt - the multiline box. Same kind of language you'd use with any modern image model; it's surprisingly good at following detail.
• resolution - 1K / 2K / 4K. This is the only thing that changes the price: 1K ≈ 39 flames, 2K ≈ 65, 4K ≈ 103 per image. The node shows an "≈ N 🔥" preview on its body and prints the estimated cost to the console.
• aspect_ratio - auto plus the usual suspects (1:1, 16:9, 9:16, 21:9, 4:3, 3:2 …). Note the provider caps: auto is 1K only, and 1:1 won't go above 2K.

Output is a single image (IMAGE), which wires straight into Save Image, Preview, or any downstream local node - so you can generate text-heavy art in the cloud and still upscale, detail, or composite it locally.

Install, as with everything in this pack: ComfyUI Manager, search "HOTCUT", install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/Smyshnikof/comfyui-hotcut.git
cd comfyui-hotcut
pip install -r requirements.txt

No model downloads, no VRAM math. The starter template is bundled - Workflow → Browse Templates → HOTCUT, or example_workflows/ in the repo.

The gotchas are the honest price of a closed model. GPT Image 2 is moderated by the provider, always - there's no toggle and no local bypass, because there are no weights to patch. If the service or moderation rejects a prompt, you're not charged (billing is strictly per successful generation), but you're also not getting an image. The node waits up to 300 seconds for a result, so a slow service day can feel like a hang. And your prompt leaves the machine - that's the mechanism, not a bug. If you need everything to stay local, this isn't the node; if you need a poster with a legible headline in Cyrillic, it's genuinely the one you'd reach for.