v1.2 (Beta 2)

API Updates:

• Add EMA support in Manager
• Add multiple devices support in training configs
• Introducing callbacks.SamplingCallback to sample images during training
• Introducing optim.EMAOptimizer for EMA optimizer with context control to swap parameters

Beta Updates:

• Compatibility fixed
• FP16 support

v1.1.1

Updates:

• Minor bugs fixed

v1.1

API Updates:

• Add forward diffusion control in train_step and test_step function
• Add SDE configurations in configs package
• Introducing build function to build non-timed wrapped UNet
• Introducing managers.Manager, one manager for all separated nn.DiffusionModule
• Introducing nn.diffusion.FastSamplingDiffusionModule
• Introducing nn.LatentMode for nn.LatentDiffusionModule to encode, decode, or forward
• Introducing sde.SDEType to load SDEs in default settings.
• Introducing the separated nn.DDPM and nn.LatentDiffusionModule
• Introducing the separated nn.DiffusionModule for separated diffusion process into PyTorch module
• Introducing the separated nn.SDEModule
• Multi-GPUs support for nn.DiffusionModule
• Parsing additional arguments when forwarding in nn.diffusion.LatentDiffusionModule
• The new separated nn.DiffusionModule no longer need a nn.TimedModule, but a normal torch.nn.Module instead for better compatibility on other unets
• Use prior_sampling in SDE instead of direct randomize when adding noises.

Other updates:

• Compatibility for non-wrapped models improved
• Minor bugs fixed
• Typing improvement

v1.1 (Release Candidate 1)

API Updates:

• Add SDE configurations in configs package
• Introducing build function to build non-timed wrapped UNet
• Introducing managers.Manager, one manager for all separated nn.DiffusionModule
• Introducing nn.diffusion.FastSamplingDiffusionModule
• Introducing nn.LatentMode for nn.LatentDiffusionModule to encode, decode, or forward
• Introducing sde.SDEType to load SDEs in default settings.
• Introducing the separated nn.DDPM and nn.LatentDiffusionModule
• Introducing the separated nn.DiffusionModule for separated diffusion process into PyTorch module
• Introducing the separated nn.SDEModule
• Multi-GPUs support for nn.DiffusionModule
• Parsing additional arguments when forwarding in nn.diffusion.LatentDiffusionModule
• The new separated nn.DiffusionModule no longer need a nn.TimedModule, but a normal torch.nn.Module instead for better compatibility on other unets
• Use prior_sampling in SDE instead of direct randomize when adding noises.

Other updates:

• Compatibility for non-wrapped models improved
• Minor bugs fixed
• Typing improvement

Release candidate updates:

• Add forward diffusion control in train_step and test_step function

v1.0.3

Updates:

• Minor bugs fixed

v1.0.2

Updates:

• Minor bugs fixed
• Remove reversed typing for sampling_range
• Replace possible type of sampling_rage from typing.Iterable[int] with typing.Sequence[int]

v1.0.1

Updates:

• Minor bugs fixed in SDE
• Remove deprecated package

v1.0

Main release:

• Introducing an abstract DiffusionManager to control all diffusion models
• Introducing data package with a noised data type data.DiffusionData and an unsupervised dataset data.SequenceDataset
• Introducing DDPMManager and SDEManager for DDPM and SDE implementation
• Intorducing metrics.LPIPS using lpips package
• Introducing scheduling package for beta scheduling in DDPM
• Introducing sde package for SDE implementation