daniel/h3-blackwell-runtime
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Decode final sampled latent in VAE gate

Browse source
This commit is contained in:
Daniel Maddern 2026-08-13 15:59:25 +07:00
parent 5062444d11
commit 15ac9b4c9c
1 changed files with 7 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

9
tools/compare_vae_pixels.py
View file

@ -21,14 +21,19 @@ def load_reference_frames(path: Path, frames: int) -> torch.Tensor:
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--latent", default="/artifacts/capture/output_11.pt")
parser.add_argument("--latent", default="/artifacts/fl2va-sampler-reference/step_11.pt")
parser.add_argument("--frames", type=int, default=22)
parser.add_argument("--reference-frames", type=Path, default=Path("/artifacts/fl2va-sampler-reference-e679178/reference_frames"))
parser.add_argument("--output", type=Path, default=Path("/output/h3-blackwell-runtime/direct-vae-same-latent.mp4"))
args = parser.parse_args()
state = torch.load(args.latent, map_location="cuda", weights_only=False)
latent = state["video"].to("cuda")
if "denoised" in state:
video_shape = (1, 24, 7, 12, 20)
video_count = torch.tensor(video_shape).prod().item()
latent = state["denoised"].to("cuda").reshape(-1)[:video_count].reshape(video_shape)
else:
latent = state["video"].to("cuda")
vae = MiniMaxH3VideoVAE.from_safetensors("/vae/minimax_h3_video_vae_fp16.safetensors", device="cuda").eval()
with torch.inference_mode():
decoded = vae.decode(latent.to(next(vae.parameters()).dtype))[:, :, : args.frames]