Skip to content

v0.8.0: Pillar-attention support

Choose a tag to compare

View all tags
@danny-mio danny-mio released this 21 Feb 23:20
· 1 commit to main since this release
Immutable release. Only release title and notes can be modified.
v0.8.0
This tag was signed with the committer’s verified signature.
danny-mio Daniele Camisani
GPG key ID: C0EB89BDE0259691
Verified
Learn about vigilant mode.
258442c
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Changed

  • Updated fluxflow dependency to >=0.8.0
  • v0.8.0 pillar-attention flow architecture supported via model_version: "0.8.0" in training config
  • FlowTrainer and pipeline unchanged; version routing handled by load_versioned_checkpoint()

Also includes (0.7.1 was not released separately)

  • VAE loss component toggles: configurable enable/disable for KL divergence, color statistics, histogram matching, contrast regularization, and coarseness losses
  • Coarseness loss: per-channel texture loss matching local patch variance distributions
  • Normalize timesteps based on active window (start_step to num_train_timesteps)
  • Correct _first_step handling so the scheduler advances properly
  • Prepare/step predictor optimizer safely with accelerator and reinit on dim changes
Assets 3
Loading