github/ai-toolkit
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Add support for audo frame count so datasets can have varrying length videos. Varous ltx 2.3 VAE optimizations such as removing tiling articacts, and doing frame split encoding to reduce vram on encoding/decoding.

Browse Source
This commit is contained in:
Jaret Burkett
2026-03-24 12:20:09 -06:00
parent 4ad14d211a
commit 7f3309b291
10 changed files with 132 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
toolkit/data_loader.py
View File

@@ -389,7 +389,7 @@ class AiToolkitDataset(LatentCachingMixin, ControlCachingMixin, CLIPCachingMixin
self.dataset_config = dataset_config
# update bucket divisibility
self.dataset_config.bucket_tolerance = sd.get_bucket_divisibility()
self.is_video = dataset_config.num_frames > 1
self.is_video = dataset_config.num_frames > 1 or dataset_config.auto_frame_count
super().__init__()
folder_path = dataset_config.folder_path
self.dataset_path = dataset_config.dataset_path
@@ -505,6 +505,13 @@ class AiToolkitDataset(LatentCachingMixin, ControlCachingMixin, CLIPCachingMixin
latent_space_version = 'sdxl'
else:
latent_space_version = self.sd.model_config.arch if self.sd is not None else "sd1"
temporal_compression = 8
if self.sd is not None:
if hasattr(self.sd.vae.config, 'scale_factor_temporal'):
temporal_compression = self.sd.vae.config.scale_factor_temporal
if hasattr(self.sd.unet.config, 'temporal_compression_ratio'):
temporal_compression = self.sd.unet.config.temporal_compression_ratio
bad_count = 0
for file in tqdm(file_list):
@@ -520,6 +527,7 @@ class AiToolkitDataset(LatentCachingMixin, ControlCachingMixin, CLIPCachingMixin
text_embedding_space_version=self.sd.model_config.arch if self.sd else "sd1",
te_padding_side=self.sd.te_padding_side if self.sd else "right",
latent_space_version=latent_space_version,
temporal_compression=temporal_compression,
)
self.file_list.append(file_item)
except Exception as e: