Superside GPT Image 2 Edit

OpenAI's GPT Image 2 is a genuinely good edit model, but the default API experience makes you think about output sizes and masks more than you should. This node fixes both of those with two well-thought-out dropdowns, and it's the version of "use GPT Image 2 in ComfyUI" that actually feels designed instead of bolted together.

The sizing story first, because it's the most confusing part of the raw API and this node solves it. GPT Image 2 caps total output at roughly 8 megapixels, so "4K" doesn't mean the same thing at every aspect ratio. The node's size dropdown defaults to match input + resolution: it keeps your input image's own aspect ratio - a tall portrait stays a tall portrait, no surprise cropping - and scales it to the resolution you pick (1K/2K/4K). So you just choose 4K and get the biggest valid output for whatever shape you fed it: ~3840px on the long edge at 16:9 (true UHD), ~2880 at 1:1, ~2528×3264 for a portrait. Other size options: match input (original) keeps the input's exact size, a fixed aspect ratio, or custom pixels (with width/height, multiples of 16). The tooltips spell all this out, and it's the rare node where the defaults are genuinely the right call.

The masking story is the second thing this node gets right. One mask_mode dropdown replaces a pile of fiddly settings:

• off - edit whole image (default) - the mask is ignored entirely. This is for crop-and-stitch pipelines where a separate stitch node does the masking, which is how the pack's own workflows treat it.
• guide model (soft) - sends the mask so GPT focuses its edits on the white area, but the model may still re-render the rest.
• lock outside mask (hard) - the same, plus the result is composited back only inside the mask, so everything outside stays pixel-identical to the input. This is the standalone-inpainting mode, and it's the one that gets you the "nothing else moved" guarantee.

invert_mask flips the white=edit convention if your mask is inverted. quality (auto/low/medium/high), num_images, and output_format (png/jpeg/webp) round out the rest.

Inputs: prompt, image_1, and api_key required; image_2–image_6 for multi-reference editing, plus mask_image and everything above. Outputs are images (IMAGE) and info (STRING, result URL).

Under the hood it runs through fal's queued execution path - GPT Image 2 is a slow endpoint, so the node submits and polls with a 20-minute timeout rather than holding one fragile connection open. Expect the graph to wait on this one.

The honest caveats: it's a closed model on metered fal credits, your images leave the machine, and OpenAI's content policy rides along - there's no local abliterate for this. The KB's read on GPT Image 1.5 is "impressive but unreliable access," and GPT Image 2 is the stabilized version of that same line. Use the lock outside mask mode when you need the rest of the frame untouched, and off mode in a stitch pipeline - the tooltips genuinely tell you which.

Install - ComfyUI Manager (search "comfyui-superside-nodes") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Superside/comfyui-superside-nodes
pip install -r requirements.txt

Restart ComfyUI, find it under Superside. Key in the api_key widget (blank falls back to FAL_KEY), no model downloads.