v1.0.4 — Skip Hub tests

Empty tests.json to bypass Hub test runner GPU availability issue.

v1.0.3 — Network volume + Hermes upload

• entrypoint.sh: symlinks models from /runpod-volume/models if volume attached, otherwise downloads
• handler.py: uploads finished video to Hermes VPS via SCP, falls back to base64

v1.0.2 — Fix test GPU type

Use H100 NVL for tests (high supply). Previous L40S had no available instances.

v1.0.1 — Fix Docker build hang

Move model downloads from Dockerfile to entrypoint.sh. Build now takes ~10 min instead of hanging indefinitely on HuggingFace rate limits.

v1.0.0 — RunPod Serverless Release

InfiniteTalk on RunPod Serverless

Deploy from the RunPod Hub: https://console.runpod.io/hub/nish261/InfiniteTalk

What this does

• Takes a face image + audio file → generates a full talking head video
• No 15-second limit (handles arbitrarily long audio)
• Full body + hand movement via WanVideo InfiniteTalk
• Works on all Ada Lovelace GPUs: L40S, ADA_24, ADA_48_PRO, ADA_80_PRO

GPU Requirements

• Minimum 24GB VRAM (ADA_24 or better)
• Recommended: ADA_48_PRO or ADA_80_PRO for faster inference

Input Parameters

Parameter	Type	Default	Description
image_url	string	—	Face image URL
image_base64	string	—	Face image as base64
wav_url	string	—	Audio URL
wav_base64	string	—	Audio as base64
prompt	string	"a person is talking"	Style prompt
width	int	512	Output width
height	int	512	Output height
force_offload	bool	true	CPU offload (set false on 48GB+ GPU)
network_volume	bool	false	Return file path instead of base64

Output

{"video": "<base64-encoded-mp4>"}