zhenzhen-creatify-aurora-fal

This is the pack's talking-avatar node. Comfly_creatify_aurora_fal takes an image of a face and an audio clip - speech, singing, whatever - and returns a video of that face performing it. It's a FAL-family node (endpoint fal-ai/creatify/aurora), which means it runs through the author's proxy with a Zhenzhen key, pay-per-call, no local compute.

Why you'd reach for it: the "make this person say this thing" job. For lip-sync and audio-driven video, the local ecosystem has options (the audio-generation KB tracks the lip-sync models), but they're fiddly, they want a GPU, and quality varies. Creatify Aurora is the closed, hosted version: image + audio in, talking video out. If you've got a product presenter, a character, or a fan-dub job, this is the turnkey path - at per-call prices.

Inputs that matter

• image_url / audio_url - both required in the schema (strings), though in practice the node accepts connected tensors too: image (IMAGE) and audio (AUDIO). The image_way / audio_way toggles pick upload method - base64 / upload (default) send connected tensors; image_url / audio_url skip the upload and use the URLs directly.
• api_key - your Zhenzhen key, in the field (no settings node for FAL nodes).
• prompt - optional; an instruction for the animation.
• guidance_scale (0–10, default 1) and audio_guidance_scale (0–10, default 2) - how hard the video follows the audio/visual conditioning. The audio dial being higher by default tells you which signal dominates.
• resolution - 480p (default) or 720p. Draft at 480p, ship at 720p.
• poll_interval / max_poll_attempts - default 6s × 600 = one-hour timeout.

Outputs: video (VIDEO), video_url (STRING), response (JSON).

Install

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

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

No model files. Zhenzhen key required - and remember the FAL channel is pre-charge with a refund after the job settles, so the balance moves at submit time.

Where people get burned

• Both inputs are genuinely required. No image, no audio, no video - the node raises "requires image/audio inputs or URLs." The image_url/audio_url fields being in required is a hint: the URL path is the designed flow, and connecting tensors is the alternative.
• A clean face makes or breaks it. The avatar quality tracks the source image. Blurry, angled, or obstructed faces produce cursed lip-sync. Feed it a clean frontal reference.
• Audio quality is the other half. The audio_guidance_scale default of 2 (double the visual's 1) is the model telling you the audio is the boss. Give it clean audio - a bad recording produces a video that confidently lip-syncs to nothing.
• 480p first. The resolution dial is a cost lever, and 480p is plenty for testing whether the animation works before you pay for 720p.
• It's a hosted filter. Faces of real people, copyright audio, and content moderation are all enforced server-side. The node can't and won't bypass that - if it refuses, the model refused.

The pitch is simple: one face + one audio clip = one talking video, no lip-sync pipeline, no GPU. It's the pack's most "just works" node for the one job it does, and the failure modes are almost entirely input hygiene - feed it a good face and good audio.