Mirrored_Repos/InvokeAI
1
0
Fork 0
mirror of https://github.com/invoke-ai/InvokeAI synced 2024-08-30 20:32:17 +00:00
Code Issues Packages Projects Releases Wiki Activity

add cross-attention support to im2img; prevent inpainting from crashing

Browse Source
This commit is contained in:
Damian at mba 2022-10-18 23:23:38 +02:00
parent 2b79a716aa
commit 582880b314
4 changed files with 15 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
ldm/generate.py
View File

@ -541,7 +541,8 @@ class Generate:
image = Image.open(image_path) image = Image.open(image_path)
# used by multiple postfixers # used by multiple postfixers
uc, c = get_uc_and_c( # todo: cross-attention
uc, c, _, _ = get_uc_and_c_and_ec(
prompt, model =self.model, prompt, model =self.model,
skip_normalize=opt.skip_normalize, skip_normalize=opt.skip_normalize,
log_tokens =opt.log_tokenization log_tokens =opt.log_tokenization

7
ldm/invoke/generator/img2img.py
View File

@ -32,7 +32,7 @@ class Img2Img(Generator):
) # move to latent space ) # move to latent space
t_enc = int(strength * steps) t_enc = int(strength * steps)
uc, c = conditioning uc, c, ec, edit_opcodes = conditioning
def make_image(x_T): def make_image(x_T):
# encode (scaled latent) # encode (scaled latent)
@ -49,7 +49,10 @@ class Img2Img(Generator):
img_callback = step_callback, img_callback = step_callback,
unconditional_guidance_scale=cfg_scale, unconditional_guidance_scale=cfg_scale,
unconditional_conditioning=uc, unconditional_conditioning=uc,
init_latent = self.init_latent, # changes how noising is performed in ksampler init_latent = self.init_latent,
edited_conditioning = ec,
conditioning_edit_opcodes = edit_opcodes
# changes how noising is performed in ksampler
) )
return self.sample_to_image(samples) return self.sample_to_image(samples)

3
ldm/invoke/generator/inpaint.py
View File

@ -45,7 +45,8 @@ class Inpaint(Img2Img):
) # move to latent space ) # move to latent space
t_enc = int(strength * steps) t_enc = int(strength * steps)
uc, c = conditioning # todo: support cross-attention control
uc, c, _, _ = conditioning
print(f">> target t_enc is {t_enc} steps") print(f">> target t_enc is {t_enc} steps")

7
ldm/models/diffusion/ksampler.py
View File

@ -132,6 +132,7 @@ class KSampler(Sampler):
use_original_steps=False, use_original_steps=False,
init_latent = None, init_latent = None,
mask = None, mask = None,
**kwargs
): ):
samples,_ = self.sample( samples,_ = self.sample(
batch_size = 1, batch_size = 1,
@ -143,7 +144,8 @@ class KSampler(Sampler):
unconditional_conditioning = unconditional_conditioning, unconditional_conditioning = unconditional_conditioning,
img_callback = img_callback, img_callback = img_callback,
x0 = init_latent, x0 = init_latent,
mask = mask mask = mask,
**kwargs
) )
return samples return samples
@ -238,6 +240,8 @@ class KSampler(Sampler):
index, index,
unconditional_guidance_scale=1.0, unconditional_guidance_scale=1.0,
unconditional_conditioning=None, unconditional_conditioning=None,
edited_conditioning=None,
conditioning_edit_opcodes=None,
**kwargs, **kwargs,
): ):
if self.model_wrap is None: if self.model_wrap is None:
@ -263,6 +267,7 @@ class KSampler(Sampler):
# so the actual formula for indexing into sigmas: # so the actual formula for indexing into sigmas:
# sigma_index = (steps-index) # sigma_index = (steps-index)
s_index = t_enc - index - 1 s_index = t_enc - index - 1
self.model_wrap.prepare_to_sample(s_index, edited_conditioning=edited_conditioning, conditioning_edit_opcodes=conditioning_edit_opcodes)
img = K.sampling.__dict__[f'_{self.schedule}']( img = K.sampling.__dict__[f'_{self.schedule}'](
self.model_wrap, self.model_wrap,
img, img,