zhenzhen-boogu-image-fal

Boogu isn't a household name the way Nano Banana or Seedream are, and that's sort of the point. It's the model you reach for when you want a capable hosted text-to-image that does a little bit of everything - text-to-image and image editing in one node, a proper size picker, safety checker control - without the queue drama of the hyped frontier models. This node wraps the fal-ai/boogu-image endpoints through the Zhenzhen API, so it's pay-per-call and runs entirely on someone else's GPU.

It's a good first hosted image node for exactly one reason: it's boring in the best way. There's no trick to it, the defaults are sane, and you can get a usable image on the third click. When you want a specific aesthetic from a specific hyped model, use that model's node. When you just need a decent image without fighting the API, this one quietly does the job.

How it works

mode picks the endpoint: text_to_image hits fal-ai/boogu-image, edit hits fal-ai/boogu-image/edit and expects an input image. Everything else is a standard Zhenzhen FAL call - POST to the proxy at ai.t8star.org/fal/..., poll every few seconds until the job completes, download, hand back an IMAGE. Nothing runs locally, and the only thing you have to supply besides a prompt is the API key.

Inputs and outputs that matter

The image_size dropdown is the star here - it has eight options including custom, and unlike some of the more opinionated models in this pack, custom actually works: pick custom and set custom_width / custom_height (step 16, up to 2048). For a quick square, square_hd is the default and fine.

• num_inference_steps - 30 default, min 20. Boogu likes a reasonable step count; don't crank it to 10 expecting speed.
• guidance_scale - 4.0 default; bump it if the model is ignoring your prompt.
• image_guidance_scale - edit mode only, 1.0 default. This is how strongly the edit stays glued to your input image. Higher = more faithful to the source, less faithful to the prompt.
• num_images - 1 to 4. Remember this is pay-per-call; four images costs four times as much.
• enable_safety_checker - on by default. If you're running commercial-ish content with strict providers, leave it on; the whole point of using a hosted model is that the filter lives upstream.
• output_format - jpeg or png. png for anything you'll edit further.
• sync_mode - off by default (async with polling); turning it on makes the request wait inline.
• seed - 0 = random, FAL caps at 65535.

Outputs are the standard trio: images (the IMAGE tensor), response (raw JSON), and image_urls (hosted URLs, one per line).

Installing it

All Zhenzhen nodes install as one pack:

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

Restart ComfyUI. Or use ComfyUI Manager and search "Comfyui-zhenzhen". No models to download - the pack's requirements are HTTP libraries, not weights.

Common issues

The usual Zhenzhen pack gotchas apply: you need a workshop account with credit, and the overseas API server can throw 443 errors from some networks (VPN with TUN mode fixes it). A 500 on first run is usually the upstream provider - re-run. If you're getting weak prompt adherence in edit mode, raise image_guidance_scale only after you've confirmed guidance_scale is in a sane range; fighting both at once is how you get mush. And when a job times out after the long default poll (600 × 6s), check the workshop site's task list before re-submitting - you might already own the result.