InvokeAI/ldm/invoke/generator/inpaint.py
Lincoln Stein 6f93dc7712 cleanup inpainting and img2img
- add a `--inpaint_replace` option that fills masked regions with
  latent noise. This allows radical changes to inpainted regions
  at the cost of losing context.
- fix up readline, arg processing and metadata writing to accommodate
  this change
- fixed bug in storage and retrieval of variations, discovered incidentally
  during testing
- update documentation
2022-10-16 08:50:55 -04:00

87 lines
3.2 KiB
Python

'''
ldm.invoke.generator.inpaint descends from ldm.invoke.generator
'''
import torch
import numpy as np
from einops import rearrange, repeat
from ldm.invoke.devices import choose_autocast
from ldm.invoke.generator.img2img import Img2Img
from ldm.models.diffusion.ddim import DDIMSampler
from ldm.models.diffusion.ksampler import KSampler
class Inpaint(Img2Img):
def __init__(self, model, precision):
self.init_latent = None
super().__init__(model, precision)
@torch.no_grad()
def get_make_image(self,prompt,sampler,steps,cfg_scale,ddim_eta,
conditioning,init_image,mask_image,strength,
step_callback=None,inpaint_replace=False,**kwargs):
"""
Returns a function returning an image derived from the prompt and
the initial image + mask. Return value depends on the seed at
the time you call it. kwargs are 'init_latent' and 'strength'
"""
# klms samplers not supported yet, so ignore previous sampler
if isinstance(sampler,KSampler):
print(
f">> Using recommended DDIM sampler for inpainting."
)
sampler = DDIMSampler(self.model, device=self.model.device)
sampler.make_schedule(
ddim_num_steps=steps, ddim_eta=ddim_eta, verbose=False
)
mask_image = mask_image[0][0].unsqueeze(0).repeat(4,1,1).unsqueeze(0)
mask_image = repeat(mask_image, '1 ... -> b ...', b=1)
scope = choose_autocast(self.precision)
with scope(self.model.device.type):
self.init_latent = self.model.get_first_stage_encoding(
self.model.encode_first_stage(init_image)
) # move to latent space
t_enc = int(strength * steps)
uc, c = conditioning
print(f">> target t_enc is {t_enc} steps")
@torch.no_grad()
def make_image(x_T):
# encode (scaled latent)
z_enc = sampler.stochastic_encode(
self.init_latent,
torch.tensor([t_enc]).to(self.model.device),
noise=x_T
)
# to replace masked area with latent noise, weighted by inpaint_replace strength
if inpaint_replace > 0.0:
print(f'>> inpaint will replace what was under the mask with a strength of {inpaint_replace}')
l_noise = self.get_noise(kwargs['width'],kwargs['height'])
inverted_mask = 1.0-mask_image # there will be 1s where the mask is
masked_region = (1.0-inpaint_replace) * inverted_mask * z_enc + inpaint_replace * inverted_mask * l_noise
z_enc = z_enc * mask_image + masked_region
# decode it
samples = sampler.decode(
z_enc,
c,
t_enc,
img_callback = step_callback,
unconditional_guidance_scale = cfg_scale,
unconditional_conditioning = uc,
mask = mask_image,
init_latent = self.init_latent
)
return self.sample_to_image(samples)
return make_image