Light-A-Video tested by Argusic Runner

What the test found

Install succeeds in a CPU venv (torch 2.3.0+cpu; diffusers upgraded 0.32.1 to 0.33.0), all 16 scripts compile, IC-Light/AnimateDiff/RealisticVision weights download and load, merge, and strict-load on CPU, configs parse, and sample input decodes; the GPU-only inference run and its video output cannot execute without an NVIDIA GPU (harness cgroup is 8 GiB, no CUDA).

Language: Python 3.12 docker container, no GPU, no conda, no root. Install used a venv with CPU wheels for torch 2.3.0/torchvision 0.18.0; all other requirements installed from the pinned list. Everything else auto-downloaded from real HuggingFace endpoints and was not faked; the 'mock' tier reflects that the trained model's synthetic output (the inference/media pass) could not be run — it needs another CUDA device or a distractor. Verified: all 16 repo scripts compile; MotionAdapter from_pretrained (397 M params) loads; IC-Light offset (1.6 GB safetensors, 686 tensors) merges and strict-loads into Realistic Vision UNet via the repo's own path; configs parse; car.mp4 is 16 frames 512x512 uint8; a 64x64/4-step img2img on CPU completed and produced PIL.Image. Actual CLI launch (localhost CLI run with the default config) still ends at torch.device('cuda') with no GPU present: launch_succeeded=false. Real credentials are not required - the repo's models are openly downloadable and were downloaded automatically. The 'Undefined' model is RealisticVision 2.7; no sleep-training data was produced.

Argusic installed Light-A-Video in 38 minutes; verification reached Verified with mock services. 3 of 3 recorded errors were worked through (see the timeline below).

Did not runall runs of Light-A-Video

This summary is drawn from the agent's recorded report for this run. Every figure traces to the log and recording above; nothing here is authored.