zhenzhen-video-g-omni-1.1-flash-lowprice

The name is honest about the one thing that matters: this node generates 4–10 second video clips and never touches your GPU. No checkpoint, no VAE, no VRAM juggling - you type a prompt, the node phones home to the zhenzhen shop's API, and a finished MP4 comes back into your graph as a normal VIDEO. It's the "lowprice" tier of their g-omni flash line, which is exactly the pattern worth understanding before you install anything: this is a paid API wrapper, not a local model.

What you're actually paying for

zhenzhen (the author, T8mars) runs a fan-service reseller shop - 贞贞的AI工坊 overseas, 贞贞的平价AI小铺 domestically - that proxies closed models at thin margins and hands you an API key. The g-omni-1.1-flash-lowprice model is a closed, Google-class omni video model (the "omni" is the tell: it ingests text, frames, reference images, or a reference video and generates video from any of them). You can't download these weights any more than you can download Veo. The reseller is how people who don't want to fight Google's billing get at them from inside ComfyUI - same calculus the KB's API-node docs describe: cost, region, billing friction. The community is right to be grumpy about API nodes as a default, but for a model you were never allowed to download, this is the door.

How it works

Synchronous and straightforward: the node validates your inputs, uploads any reference media (images as PNGs, video as MP4) to the shop, POSTs a job, then polls it every ~4 seconds until it finishes (up to 30 minutes), downloads the result, and hands it back. You watch the progress bar crawl. What comes out:

• video - the clip as a ComfyUI VIDEO, ready for a SaveVideo node
• video_url - the hosted MP4 link
• task_id - your job's ID, handy if the node times out and you need to pull the clip from the shop's async-task console
• response - the full JSON, which is where you'll actually find the error message when something breaks

The few inputs that matter

The whole mode system is mode plus whatever it demands:

• text - plain text-to-video, no reference media allowed
• frame - exactly one image as the first frame
• reference_images - one or three images, contiguous from image1
• reference_video - exactly one input_video or a video_url

Then seconds (4/6/8/10), resolution (720p/1080p/4k), aspect_ratio (16:9/9:16), and nsfw_check. The api_config socket comes from the pack's Zhenzhen API Settings node (Comfly_api_set, set api_base to seedance_low_price and paste your key). skip_error makes failures return a blank/error video instead of throwing, and seed is cache-control only - the tooltip is explicit that it's not sent to the API, so don't expect to reproduce a clip by locking a seed.

Installing it

No model downloads, no weights. Install the pack, restart, paste a key.

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

or search Comfyui-zhenzhen in ComfyUI Manager and let it handle the requirements.txt (aiohttp, requests, opencv-python, transformers, openai - the usual suspects). Then grab an API key from api.seedance.nz (domestic shop) or ai.t8star.org (overseas). The pack ships ready-made workflows for all four modes in its workflow/ folder, and they deliberately don't save your key.

Where people get burned

• 500 errors. Mostly the provider, not you. The README's standing advice: run it again - cheap-clip services have flaky moments.
• 443 / "can't connect" errors. The shop runs on overseas servers; the README says plainly you need a VPN with Tun mode, and to kill 360-style firewalls that block it.
• Strict mode validation. These nodes reject sloppy wiring at the front door - two reference images when the model wants one or three, an image in text mode, etc. The error message tells you exactly what it wants.
• Registration. The overseas shop has periodically closed new sign-ups; the domestic one stays open. Check before you plan a workflow around it.

Per-call cost adds up fast if you're sloppy - but for one-off clips of a closed model you can't run locally, it's hard to beat. Want to batch ten of these without ComfyUI choking? That's the concurrent submit variant - same node, split in two.