add dpm-solver support (much faster than plms)

[LuChengTHU]

DPM-Solver (https://github.com/LuChengTHU/dpm-solver) is much faster than PLMS and DDIM, which can generate almost converged samples in only 20-25 steps.

Example images of the difference between DDIM, PLMS, and DPM-Solver can be found in: https://github.com/LuChengTHU/dpm-solver/tree/main/example_v2/stable-diffusion

[LuChengTHU]

An example command for sampling by DPM-Solver:

python scripts/txt2img.py --prompt "a photograph of an astronaut riding a horse" --dpm_solver --ddim_steps 25

[rromb]

This is great, thank you very much 👍

[rajbala]

I used the instructions to install the dependencies:

conda env create -f environment.yaml
conda activate ldm

But I get the following error when I try to run python scripts/txt2img.py --prompt "a photograph of an astronaut riding a horse"

[Traceback (most recent call last):
  File "scripts/txt2img.py", line 20, in <module>
    from ldm.models.diffusion.dpm_solver import DPMSolverSampler
  File "/home/opc/stable-diffusion/ldm/models/diffusion/dpm_solver/__init__.py", line 1, in <module>
    from .sampler import DPMSolverSampler
  File "/home/opc/stable-diffusion/ldm/models/diffusion/dpm_solver/sampler.py", line 5, in <module>
    from .solver import NoiseScheduleVP, model_wrapper, DPM_Solver
ModuleNotFoundError: No module named 'ldm.models.diffusion.dpm_solver.solver'
]

Any ideas as to what is happening?

[rromb]

should be fixed in 21f890f

[rajbala]

It is fixed indeed! :)