GPT-Image 1.5 Edit (fal)

This is OpenAI's image model - the one behind ChatGPT's image generation - reachable from a ComfyUI graph. GPTImage15Edit_fal calls GPT-Image 1.5's edit endpoint on fal: you feed it one or more images and a prompt, and it returns an edited IMAGE. GPT-Image is closed and API-only, so an API node is the only way it shows up in ComfyUI at all. It's the model to reach for when you want its particular strengths - strong instruction-following, clean in-image text, and genuinely good transparent-background output.

Where it stands out from the other editors in this pack is transparent backgrounds and native masking. You can ask for a cutout-ready PNG, or hand it a mask to constrain the edit to a region. That makes it a real option for product shots, stickers, and compositing work, not just general "change this to that" editing.

The inputs that matter

• prompt - the instruction, plain language. GPT-Image follows detailed prompts well, so be specific.
• images - the source image(s) to edit.
• mask_image - optional; supply it to limit the edit to a masked region (inpaint-style) instead of the whole frame.
• background - auto, transparent, or opaque. Set it to transparent when you want a cutout.
• input_fidelity - low or high; high preserves more of the source's original detail, which you usually want for edits.

Rounding it out: image_size (auto or fixed dims), quality (low/medium/high), num_images (up to 4), output_format (jpeg/png/webp - pick png or webp for transparency), and sync_mode. One IMAGE output.

Installing it

ComfyUI Manager → search ComfyUI-fal-API → install → restart. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/gokayfem/ComfyUI-fal-API.git
pip install -r ComfyUI-fal-API/requirements.txt

No models to download - GPT-Image's weights are OpenAI's, hosted, never local. Set your fal key (config.ini, or export FAL_KEY=your_key; fal.ai/dashboard/keys), restart, find it under FAL.

Where people get burned

Two things. First, the format/background mismatch: asking for a transparent background while output_format is jpeg gets you no transparency, because JPEG can't store an alpha channel. Pair background: transparent with png or webp. Second, GPT-Image runs OpenAI's content policy, which is stricter than most - edits it considers off-limits (certain people, certain content) come back refused. The pack surfaces that as fal's real error message, so read it rather than assuming the node broke.

It's also on the pricier end per call among the image editors, so lean on the pack's result cache while you dial in a prompt, and reach for SeedEditV3_fal or the Qwen edit node when you want a cheaper or more permissive editor. Like all instruction editors, it re-generates the frame - use the mask_image input when you need the rest of the picture to stay put.