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

Merge branch 'main' into lstein/config-management-fixes

Browse Source
This commit is contained in:
Lincoln Stein
2023-05-29 21:03:03 -04:00
committed by GitHub
parent 951900a86a 070218aba7
commit 10fe31c2a1
161 changed files with 3736 additions and 1632 deletions
Download Patch File Download Diff File Expand all files Collapse all files

34
invokeai/backend/stable_diffusion/diffusion/shared_invokeai_diffusion.py
View File

@ -183,6 +183,7 @@ class InvokeAIDiffuserComponent:
unconditional_guidance_scale: float,
step_index: Optional[int] = None,
total_step_count: Optional[int] = None,
**kwargs,
):
"""
:param x: current latents
@ -211,7 +212,7 @@ class InvokeAIDiffuserComponent:
if wants_hybrid_conditioning:
unconditioned_next_x, conditioned_next_x = self._apply_hybrid_conditioning(
x, sigma, unconditioning, conditioning
x, sigma, unconditioning, conditioning, **kwargs,
)
elif wants_cross_attention_control:
(
@ -223,13 +224,14 @@ class InvokeAIDiffuserComponent:
unconditioning,
conditioning,
cross_attention_control_types_to_do,
**kwargs,
)
elif self.sequential_guidance:
(
unconditioned_next_x,
conditioned_next_x,
) = self._apply_standard_conditioning_sequentially(
x, sigma, unconditioning, conditioning
x, sigma, unconditioning, conditioning, **kwargs,
)
else:
@ -237,7 +239,7 @@ class InvokeAIDiffuserComponent:
unconditioned_next_x,
conditioned_next_x,
) = self._apply_standard_conditioning(
x, sigma, unconditioning, conditioning
x, sigma, unconditioning, conditioning, **kwargs,
)
combined_next_x = self._combine(
@ -284,13 +286,13 @@ class InvokeAIDiffuserComponent:
# methods below are called from do_diffusion_step and should be considered private to this class.
def _apply_standard_conditioning(self, x, sigma, unconditioning, conditioning):
def _apply_standard_conditioning(self, x, sigma, unconditioning, conditioning, **kwargs):
# fast batched path
x_twice = torch.cat([x] * 2)
sigma_twice = torch.cat([sigma] * 2)
both_conditionings = torch.cat([unconditioning, conditioning])
both_results = self.model_forward_callback(
x_twice, sigma_twice, both_conditionings
x_twice, sigma_twice, both_conditionings, **kwargs,
)
unconditioned_next_x, conditioned_next_x = both_results.chunk(2)
if conditioned_next_x.device.type == "mps":
@ -304,16 +306,17 @@ class InvokeAIDiffuserComponent:
sigma,
unconditioning: torch.Tensor,
conditioning: torch.Tensor,
**kwargs,
):
# low-memory sequential path
unconditioned_next_x = self.model_forward_callback(x, sigma, unconditioning)
conditioned_next_x = self.model_forward_callback(x, sigma, conditioning)
unconditioned_next_x = self.model_forward_callback(x, sigma, unconditioning, **kwargs)
conditioned_next_x = self.model_forward_callback(x, sigma, conditioning, **kwargs)
if conditioned_next_x.device.type == "mps":
# prevent a result filled with zeros. seems to be a torch bug.
conditioned_next_x = conditioned_next_x.clone()
return unconditioned_next_x, conditioned_next_x
def _apply_hybrid_conditioning(self, x, sigma, unconditioning, conditioning):
def _apply_hybrid_conditioning(self, x, sigma, unconditioning, conditioning, **kwargs):
assert isinstance(conditioning, dict)
assert isinstance(unconditioning, dict)
x_twice = torch.cat([x] * 2)
@ -328,7 +331,7 @@ class InvokeAIDiffuserComponent:
else:
both_conditionings[k] = torch.cat([unconditioning[k], conditioning[k]])
unconditioned_next_x, conditioned_next_x = self.model_forward_callback(
x_twice, sigma_twice, both_conditionings
x_twice, sigma_twice, both_conditionings, **kwargs,
).chunk(2)
return unconditioned_next_x, conditioned_next_x
@ -339,6 +342,7 @@ class InvokeAIDiffuserComponent:
unconditioning,
conditioning,
cross_attention_control_types_to_do,
**kwargs,
):
if self.is_running_diffusers:
return self._apply_cross_attention_controlled_conditioning__diffusers(
@ -347,6 +351,7 @@ class InvokeAIDiffuserComponent:
unconditioning,
conditioning,
cross_attention_control_types_to_do,
**kwargs,
)
else:
return self._apply_cross_attention_controlled_conditioning__compvis(
@ -355,6 +360,7 @@ class InvokeAIDiffuserComponent:
unconditioning,
conditioning,
cross_attention_control_types_to_do,
**kwargs,
)
def _apply_cross_attention_controlled_conditioning__diffusers(
@ -364,6 +370,7 @@ class InvokeAIDiffuserComponent:
unconditioning,
conditioning,
cross_attention_control_types_to_do,
**kwargs,
):
context: Context = self.cross_attention_control_context
@ -379,6 +386,7 @@ class InvokeAIDiffuserComponent:
sigma,
unconditioning,
{"swap_cross_attn_context": cross_attn_processor_context},
**kwargs,
)
# do requested cross attention types for conditioning (positive prompt)
@ -390,6 +398,7 @@ class InvokeAIDiffuserComponent:
sigma,
conditioning,
{"swap_cross_attn_context": cross_attn_processor_context},
**kwargs,
)
return unconditioned_next_x, conditioned_next_x
@ -400,6 +409,7 @@ class InvokeAIDiffuserComponent:
unconditioning,
conditioning,
cross_attention_control_types_to_do,
**kwargs,
):
# print('pct', percent_through, ': doing cross attention control on', cross_attention_control_types_to_do)
# slower non-batched path (20% slower on mac MPS)
@ -413,13 +423,13 @@ class InvokeAIDiffuserComponent:
context: Context = self.cross_attention_control_context
try:
unconditioned_next_x = self.model_forward_callback(x, sigma, unconditioning)
unconditioned_next_x = self.model_forward_callback(x, sigma, unconditioning, **kwargs)
# process x using the original prompt, saving the attention maps
# print("saving attention maps for", cross_attention_control_types_to_do)
for ca_type in cross_attention_control_types_to_do:
context.request_save_attention_maps(ca_type)
_ = self.model_forward_callback(x, sigma, conditioning)
_ = self.model_forward_callback(x, sigma, conditioning, **kwargs,)
context.clear_requests(cleanup=False)
# process x again, using the saved attention maps to control where self.edited_conditioning will be applied
@ -430,7 +440,7 @@ class InvokeAIDiffuserComponent:
self.conditioning.cross_attention_control_args.edited_conditioning
)
conditioned_next_x = self.model_forward_callback(
x, sigma, edited_conditioning
x, sigma, edited_conditioning, **kwargs,
)
context.clear_requests(cleanup=True)