zhenzhen-gpt-image-2-official_ratio

The official_ratio node is the one you want when gpt-image-2 needs to come out at a specific aspect ratio - including the weird ones. Thirteen ratio choices from auto through the cinematic 21:9 and portrait 9:16, three resolutions (1k/2k/4k), and up to sixteen reference images plus an optional MASK for editing. If you're building a consistent visual asset set - a product line, a storyboard, a social-media batch where every image has to fit a specific crop - this is the node.

That "sixteen inputs" number is the standout. Most gpt-image-2 wrappers stop at four or five; this one takes a whole folder's worth and lets the model reason across them. The changelog shows the pack author bumping it from 5 to 16 and adding skip_error, so it's been shaped by real usage.

The inputs that matter

• prompt - the instruction, plus aspect_ratio and resolution - the required core.
• image1–image16 and mask - the reference/edit material. With a mask (MASK tensor) it's a localized edit; without, whole-image.
• model - gpt-image-2, gpt-image-2-all, gpt-image-2-2K, or gpt-image-2-4K. The -2K/-4K variants are the pack's dedicated high-res lanes, added as stable fallbacks.
• n - 1 to 10 images per call. Meter runs per image.
• quality, background (auto/opaque), output_format, output_compression (0–100), moderation.
• response_format - url or b64_json; the README suggests url if you're sending many reference images (lighter payload).
• async_mode (default true) - goes through the relay's async task queue, with poll_interval, max_poll_attempts, max_retries, and initial_timeout controlling the wait. webhook if you want a callback instead.

Outputs: image (IMAGE tensor), image_url, and response.

Install and key

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-zhenzhen

Restart, or ComfyUI Manager → "Comfyui-zhenzhen". Key from ai.t8star.org.

The part the README yells about

This is an official-route node, and the official route cares about your token group. The README is explicit: for 2K/4K stability you want the openai优质 premium group on your token; the default group silently walks the cheaper reverse-engineered lane and will give you slow or failed high-res jobs. Get the group right on the website before you blame the node. And since async is on by default, remember the timeout rule the pack documents everywhere: if ComfyUI gives up, the job is often still cooking on the server - check the site's async task list and pull the result from there rather than paying twice. Sixteen reference images is a lot of context leaving your machine; treat the node accordingly.