3f04a0aea2
Each spatial tile decoded by MiniMaxH3VideoVAE.tiled_decode was normalized to its own local -1..1 rotary position range instead of its true offset within the full latent grid, since tiling defaults to on (tile_size=256) at any resolution above that. The ViT3D decoder then saw every tile as if it were the whole frame, so adjacent tiles came out with different learned appearance and visible seams (#15548). create_token_ids now takes the full latent grid size and the tile's offset within it, and tiled_decode passes them through so tile position ids match a single-shot decode of the whole frame. |
||
|---|---|---|
| .. | ||
| folder_path_test.py | ||
| model_detection_test.py | ||
| seedvr_vae_forward_test.py | ||
| test_minimax_h3_vae_tiled_decode.py | ||
| test_seedvr2_dtype.py | ||
| test_seedvr2_internals.py | ||
| test_seedvr2_model.py | ||
| test_seedvr2_vae_decode.py | ||
| test_seedvr2_vae_tiled.py | ||
| test_vae_decode_tiled_nested.py | ||