Azula 0.4.0

✨ What's new

• New Tweedie Moment Projected Diffusion (TMPD) guidance (#17)
• New Twisted Diffusion Sampling (TDS) guidance (31c33e9)
• New Diffusion Plug-and-Play Image Restoration (DiffPIR) guidance (3da9fd5)
• New RePaint guidance (46f714d)
• New Pseudo-inverse Guided Diffusion Model (PGDM) guidance (613a6dd)
• Updated guidance tutorial (21e2c7f)
• Refactored plugins API (7cd0121)
  Wrappers around pre-trained models have been standardized to the same formalism, inspired by Karras et al. (2022). Each wrapper defines the coefficients $c_\text{skip}$, $c_\text{out}$, $c_\text{in}$, and $c_\text{time}$ such that the pre-trained backbone $b_\phi$ parameterizes a denoiser

$$ d_\phi(x_t, t) = c_\text{skip}(t) x_t + c_\text{out}(t) b_\phi(c_\text{in}(t) x_t, c_\text{time}(t)) $$

• New Flux plugin (dc90f4d)
• New Sana plugin (790351a)
• New Stable Diffusion plugin (950494b)
• Updated latent diffusion tutorial (2c39f2a)

Full Changelog: 0.3.0...0.4.0

Azula 0.3.0

First official release of Azula 🥳

✨ What's new

• New Predictor-Corrector (PC) sampler (6990eec)
• New Adams-Bashforth (AB) and exponential Adams-Bashforth (EAB) samplers (51fac6f)
  1. The AB sampler is equivalent to the linear multi-step (LMS) sampler from crowsonkb/k-diffusion.
  2. The EAB sampler is a multi-step generalization of the DPM-Solver sampler from LuChengTHU/dpm-solver
• New cosine and rectified noise schedules (9d48dc2)
• New Diffusion Posterior Sampling (DPS) guidance method (4e793d5)
• New Moment Matching Posterior Sampling (MMPS) guidance method (ec2e858)
• New VDM plugin (#10)
• New EDM plugin (#9)
• New EDM2 plugin (#14)
• New tutorials (7e9352e, ec2102e)
• Support for any-dimensional tensors (62240c0)
• Improved U-Net architecture (2d727cd)

Full Changelog: 0.0.0...0.3.0