Zhenzhen GPT Image 2 Fal

gpt-image-2 is OpenAI's current flagship image model, and this node reaches it through the fal route - which, on this pack, is the "fast and stable" lane. The README brags and it's mostly earned: 2K and 4K output, 4K in around 70 seconds, effectively unlimited concurrency, and custom resolution support. The catch is the one line buried in the changelog: "注意设置的自定义尺寸一定要被16整除" - your custom width and height must be divisible by 16, or the request fails. Write that down.

Comfly_gpt_image_2_fal does both generate and edit (the mode toggle), and being the fal route it's pre-deduction billing: the relay holds a base amount, then settles the real cost after the job based on quality, input images, and output size.

The inputs that matter

• prompt and mode (generate or edit) - the two required ones.
• image1–image4 and mask - the edit material when mode is edit. Note the mask here is an IMAGE-typed input (a mask image), not a MASK tensor.
• image_size - auto, the preset ratios (square_hd, square, portrait_4_3, portrait_16_9, landscape_4_3, landscape_16_9), or custom. On custom, custom_width and custom_height kick in - up to 3840×2160, and again: both must be multiples of 16.
• quality - auto, high, medium, low.
• num_images - 1 to 4.
• output_format - png, jpeg, webp.
• image_way - image_url (default) or base64 for how your inputs travel.
• seed - capped at FAL's 65535 (0 = random). poll_interval and max_poll_attempts control the wait loop; the default 600 × 6s is a full hour of patience.

Outputs: images (IMAGE tensor), response (JSON), image_urls (hosted results).

Install and key

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

Restart, or ComfyUI Manager → "Comfyui-zhenzhen". The fal nodes use the same Zhenzhen key - ai.t8star.org - no separate fal.ai account needed; the relay fronts the fal endpoint. Default token group is fine for this lane.

Reality check

The fal lane is the one people actually report as stable for 4K, so if the official-route nodes are giving you white images or timeouts, this is the fallback to try. Watch the billing model: fal is pre-deducted and settles after, so the deduction line isn't the final price - the README flags this explicitly. And as always with this pack: 500s are usually upstream, re-run before debugging, and your prompts and reference images are leaving the machine. The 16-divisibility rule is the one gotcha unique to this node - a 1280×720 custom size is fine, a 1279×720 will waste a credit on an error.