FL GPT Image1 ADV

The batch-editing workhorse for OpenAI's gpt-image-1 - up to 100 separate prompt-and-image pairs, each processed concurrently, one shared settings block (size, quality, background, format) applied across all of them. If you've got a stack of images that each need a genuinely different edit instruction, not the same prompt applied to all of them, this is built for exactly that.

The mechanism

inputcount (1-100) is the dynamic-socket control - raise it and the node exposes more prompt_N/image_N pairs (prompt_1/image_1 is the first, shown by default). Each pair fires its own async call to OpenAI's gpt-image-1 endpoint, running concurrently through a thread pool rather than one after another, with results combined into a single batched IMAGE tensor plus a per-call log of API responses.

Inputs and outputs that matter

api_key is a required on-node field - unlike FL_Dalle3 and FL_GPT_Vision, this node exposes its own key widget, though the pack's own description notes this family falls back to the OPENAI_API_KEY environment variable if left blank. inputcount sets how many prompt/image pairs you're actually using. prompt_1 defaults to "Describe image 1," a placeholder worth replacing per slot. image_1 (optional) can be left empty for pure generation from that slot's prompt, or filled in for an edit of that specific image.

The shared settings apply identically across every slot: size_setting (three fixed sizes - 1024x1024, 1536x1024, 1024x1536; gpt-image-1 doesn't do arbitrary dimensions), quality_setting (auto/high/medium/low, with high costing the most per image), background_setting (auto/transparent/opaque - transparent is the one worth knowing, it's how you get an actual alpha-channel result instead of a flat background), and output_format_setting (png/jpeg/webp). seed_setting is optional.

Outputs: images (the combined batch across every slot) and API_responses (a detailed per-call log - genuinely useful for figuring out which specific slot in a 20-image batch failed or got refused, rather than guessing from a blank result).

Installing it

ComfyUI Manager: search Fill-Nodes, install, restart. Manually: cd ComfyUI/custom_nodes && git clone https://github.com/filliptm/ComfyUI_Fill-Nodes, then restart.

Where people get burned

At high inputcount this is genuinely a lot of simultaneous API calls hitting OpenAI's rate limits, not just a lot of money - if calls start failing partway through a large batch, check API_responses before assuming the node itself is broken; it's very likely a rate-limit or content-policy rejection on a specific slot, logged right there.

And background_setting=transparent only does anything useful if output_format_setting is png or webp - transparency doesn't survive a jpeg round-trip, so pairing transparent with jpeg output can't actually produce what it sounds like it should.