HOTCUT · GPT Image 2 Edit

The same model, pointed the other direction. HotcutGptImage2Edit takes an existing IMAGE, sends it to the cloud with your prompt, and returns the edited version - so you can fix, redraw, restyle or add text to something you already made, and keep the whole thing inside one ComfyUI graph.

Mechanically it's your prompt plus a round trip. The node serializes your image tensor to PNG, uploads it to POST /api/v1/uploads, then submits a task with mode: image-to-image and the returned URL, polls for the result, and decodes it back to an IMAGE. One quirk worth knowing: if you feed it a batch, it takes the first frame only (image[0]) - this is a single-image editor, not a batch node.

Inputs: config, image (IMAGE), prompt (multiline), resolution and aspect_ratio - same 1K/2K/4K choices and the same 11 ratios as the text-to-image node. Pricing is identical because it tracks resolution alone: 1K ≈ 39, 2K ≈ 65, 4K ≈ 103 flames, with the same provider caps (auto = 1K only, 1:1 max 2K). The upload adds a bit of latency on top of the generation, but no extra charge.

Where it shines is the hybrid workflow: generate a clean render locally (or with the sibling GPT Image 2 node), run this to change a sign's text, fix a mangled hand, restyle the background, or add Cyrillic lettering - then pipe the image output back into your local post-processing. That "generate locally, fix in the cloud" loop is exactly what the bundled Всё в одном (all-in-one) template chains together. The Edit node is the piece that makes the pack feel like a real part of your graph rather than a web app with extra steps.

Same install story as the rest of the pack: ComfyUI Manager → search "HOTCUT" → install → restart, or clone manually:

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

Nothing to download, nothing local runs.

The same caveats apply as to the text node: the provider moderates everything, there's no bypass, and you're only charged on success. Two edit-specific things to expect. Editing is less deterministic than generating - "change the text to X" is reliable, "make this exactly right" usually takes a couple of attempts, so budget the flames for retries. And because your image uploads to the service, don't feed it anything you wouldn't send to a cloud API. If that's a dealbreaker, the local alternative for this job is inpainting with your own checkpoint - but for the text-and-faces cases, this is the node that just works.