zhenzhen-sora2-fal

Where the pack's other Sora nodes use the chat/OpenAI-style endpoints, Comfly_sora2_fal routes Sora 2 through the fal queue - and that buys you a few things the others don't: a character_ids field for reusing a persistent character, image-to-video as a first-class mode, and finer duration control (4–20 seconds in steps of 4). It's the "I want Sora but I also want it in my fal-style workflow with character persistence" node.

Like every FAL node here, the token goes in the node's own api_key field and traffic runs through zhenzhen's proxy (https://ai.t8star.org/fal). It submits to fal-ai/sora-2/text-to-video or fal-ai/sora-2/image-to-video depending on your mode, polls, and returns a real VIDEO tensor.

The inputs that matter

• prompt - required.
• mode - auto (default), text_to_video, or image_to_video. auto picks image-to-video when an image is connected, otherwise text.
• image / image_url - reference frame for image-to-video. If image is connected it wins; otherwise image_url (a public URL) is used.
• model - sora-2 (default), sora-2-2025-12-08, or sora-2-2025-10-06 - dated snapshots, so you can pin the version you got good results from.
• duration - 4, 8, 12, 16, or 20 seconds (4 default - cheap).
• aspect_ratio - auto, 16:9, or 9:16. Note auto is only truly automatic for image-to-video; text-to-video sends 16:9.
• resolution - auto (→720p for text-to-video) or 720p.
• character_ids - the interesting one: up to two persistent character IDs, comma/newline separated. This ties into the pack's Sora character system - generate a character once, reference it across videos.
• delete_video - on by default; removes the generated video from the provider after download (good for privacy, and the README says the pack added this by default to prevent leaks).
• detect_and_block_ip - off by default; a provider-side safety switch.

Outputs: video, video_url, response. Standard fal polling knobs (poll_interval 6s, max_poll_attempts 600 ≈ 1h).

Installing it

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

or ComfyUI Manager (search "Comfyui-zhenzhen"), restart, paste a zhenzhen token (overseas https://ai.t8star.org) into api_key.

Common issues

• 500s - retry. Same rule as the whole Sora line: most failures are OpenAI-side, and the README's standing advice is to just run it again.
• Character persistence doesn't stick - character_ids has to match IDs the pack's character system actually created; re-check how you generated the character first.
• Aspect ratio surprises - if you picked auto on text-to-video and got 16:9, that's expected behavior, not a bug.

Sora's the finicky one in this pack, and the README is honest that even the author gets burned by OpenAI-side flakiness. But if you want Sora with character reuse and an image-to-video door, this node has the most knobs of the three.