zhenzhen-cosmos-3-super-fal

NVIDIA's Cosmos 3 Super - the world-model line aimed at robotics and synthetic data - squashed into one node with a mode switch that flips between text-to-image and image-to-video. This is a FAL-family node: the pack routes it through the author's proxy at {base_url}/fal/..., you drop a Zhenzhen key into the api_key field, and the heavy lifting happens on NVIDIA's servers via fal.ai.

If you're a ComfyUI regular used to local checkpoints, this is a different animal entirely: no weights, no VRAM, pay-per-call, and the output comes back as a video tensor that drops straight into your graph as if a local sampler made it. The FAL channel is pre-charge - the README is explicit that FAL models pre-deduct a credit balance and refund the difference after the job settles. Your key's balance moves in both directions.

The two modes

• text_to_image (default) - prompt → images. Controls: image_size (from square up through portrait/landscape ratios plus a custom_832x480 option), num_images (1–4), num_inference_steps (default 28), guidance_scale (default 4), negative_prompt, output_format (jpeg/png).
• image_to_video - prompt + an image → a video clip. Adds num_frames (25–189, default 49), frames_per_second (8–30, default 24). Requires the image input or an image_url.

Both modes expose the Cosmos extras: enable_prompt_expansion, enable_agentic_generation, and enable_safety_checker (default on - Cosmos ships a real safety checker, and it's a good default).

Inputs that matter

• prompt - required, multiline. The whole creative input.
• mode - text_to_image / image_to_video.
• api_key - your Zhenzhen key, typed in the field (no settings node for FAL nodes).
• seed - 0–65535. The tooltip spells it out: 0 = random, and FAL's seed caps at 65535. This is a real sampler seed, not the cache-control fake of the concurrent family.
• image / image_url + image_way - base64 (default) or image_url; the image tensor path uploads as base64, the URL path skips the upload.
• poll_interval / max_poll_attempts - polling cadence, default 6s × 600 = a 3600s (one-hour) timeout. Video jobs run long; this is the patience dial.

Outputs: images (IMAGE), video (VIDEO), response (JSON string), url (the media URL).

Install

ComfyUI Manager → search "Comfyui-zhenzhen" → install → restart, or:

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

No model downloads. Needs a Zhenzhen key.

Where people get burned

• The key goes in the node. There's no api_config input here - FAL nodes take a plain apikey widget. Wrong config pattern is the #1 stumble coming from the Seedance2-family nodes.
• Pre-charge, not pay-at-end. The FAL channel deducts up front and refunds after. A failed job shouldn't cost you, but the balance moves immediately and settles later - don't panic at the initial deduction.
• Seed max 65535. Typing 123456789 just gets clamped. If you're hunting reproducible results, stay under the cap.
• image_to_video needs an image. No image, no URL, and the mode raises "requires image or image_url" - that's validation working as intended.
• Long polls are normal. A 189-frame job at 30fps is ~6 seconds of video that took minutes to render. The 600-attempt default exists because video genuinely takes that long; lowering it just buys you false timeouts.
• It's still a reseller lane. Prompts and reference images leave your machine for a third-party proxy and then NVIDIA's servers. Fine for a closed model - just know it's the deal.

Cosmos 3 Super is a world-model tier that most ComfyUI users don't touch. If you're here, it's because you want NVIDIA-grade synthetic data or image-to-video from a video-family model without owning the hardware. The node makes that a prompt, a mode, and a key away.