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

Use textual inversion manager from pipeline, remove extra conditioning info for uc

Browse Source
This commit is contained in:
StAlKeR7779 2023-04-25 12:53:13 +03:00
parent 8cb2fa8600
commit 37916a22ad
2 changed files with 35 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

55
invokeai/app/invocations/compel.py
View File

@ -75,52 +75,63 @@ class CompelInvocation(BaseInvocation):
# TODO: global? input? # TODO: global? input?
#use_full_precision = precision == "float32" or precision == "autocast" #use_full_precision = precision == "float32" or precision == "autocast"
use_full_precision = False #use_full_precision = False
textual_inversion_manager = TextualInversionManager( # TODO: redo TI when separate model loding implemented
tokenizer=tokenizer, #textual_inversion_manager = TextualInversionManager(
text_encoder=text_encoder, # tokenizer=tokenizer,
full_precision=use_full_precision, # text_encoder=text_encoder,
# full_precision=use_full_precision,
#)
def load_huggingface_concepts(concepts: list[str]):
pipeline.textual_inversion_manager.load_huggingface_concepts(concepts)
# apply the concepts library to the prompt
positive_prompt_str = pipeline.textual_inversion_manager.hf_concepts_library.replace_concepts_with_triggers(
self.positive_prompt,
lambda concepts: load_huggingface_concepts(concepts),
pipeline.textual_inversion_manager.get_all_trigger_strings(),
)
negative_prompt_str = pipeline.textual_inversion_manager.hf_concepts_library.replace_concepts_with_triggers(
self.negative_prompt,
lambda concepts: load_huggingface_concepts(concepts),
pipeline.textual_inversion_manager.get_all_trigger_strings(),
) )
# lazy-load any deferred textual inversions. # lazy-load any deferred textual inversions.
# this might take a couple of seconds the first time a textual inversion is used. # this might take a couple of seconds the first time a textual inversion is used.
textual_inversion_manager.create_deferred_token_ids_for_any_trigger_terms( pipeline.textual_inversion_manager.create_deferred_token_ids_for_any_trigger_terms(
self.positive_prompt + "[" + self.negative_prompt + "]" positive_prompt_str + "[" + negative_prompt_str + "]"
) )
compel = Compel( compel = Compel(
tokenizer=tokenizer, tokenizer=tokenizer,
text_encoder=text_encoder, text_encoder=text_encoder,
textual_inversion_manager=textual_inversion_manager, textual_inversion_manager=pipeline.textual_inversion_manager,
dtype_for_device_getter=torch_dtype, dtype_for_device_getter=torch_dtype,
truncate_long_prompts=self.truncate_long_prompts, truncate_long_prompts=self.truncate_long_prompts,
) )
# TODO: support legacy blend? # TODO: support legacy blend?
positive_prompt: Union[FlattenedPrompt, Blend] = Compel.parse_prompt_string(self.positive_prompt) positive_prompt: Union[FlattenedPrompt, Blend] = Compel.parse_prompt_string(positive_prompt_str)
negative_prompt: Union[FlattenedPrompt, Blend] = Compel.parse_prompt_string(self.negative_prompt) negative_prompt: Union[FlattenedPrompt, Blend] = Compel.parse_prompt_string(negative_prompt_str)
if getattr(Globals, "log_tokenization", False): if getattr(Globals, "log_tokenization", False):
log_tokenization(positive_prompt, negative_prompt, tokenizer=tokenizer) log_tokenization(positive_prompt, negative_prompt, tokenizer=tokenizer)
# TODO: add lora(with model and clip field types) # TODO: add lora(with model and clip field types)
c, c_options = compel.build_conditioning_tensor_for_prompt_object(positive_prompt) c, options = compel.build_conditioning_tensor_for_prompt_object(positive_prompt)
uc, uc_options = compel.build_conditioning_tensor_for_prompt_object(negative_prompt) uc, _ = compel.build_conditioning_tensor_for_prompt_object(negative_prompt)
if not self.truncate_long_prompts: if not self.truncate_long_prompts:
[c, uc] = compel.pad_conditioning_tensors_to_same_length([c, uc]) [c, uc] = compel.pad_conditioning_tensors_to_same_length([c, uc])
c_ec = InvokeAIDiffuserComponent.ExtraConditioningInfo( ec = InvokeAIDiffuserComponent.ExtraConditioningInfo(
tokens_count_including_eos_bos=get_max_token_count(tokenizer, positive_prompt), tokens_count_including_eos_bos=get_max_token_count(tokenizer, positive_prompt),
cross_attention_control_args=c_options.get("cross_attention_control", None), cross_attention_control_args=options.get("cross_attention_control", None),
)
uc_ec = InvokeAIDiffuserComponent.ExtraConditioningInfo(
tokens_count_including_eos_bos=get_max_token_count(tokenizer, negative_prompt),
cross_attention_control_args=uc_options.get("cross_attention_control", None),
) )
name_prefix = f'{context.graph_execution_state_id}__{self.id}' name_prefix = f'{context.graph_execution_state_id}__{self.id}'
@ -128,8 +139,8 @@ class CompelInvocation(BaseInvocation):
name_negative = f"{name_prefix}_negative" name_negative = f"{name_prefix}_negative"
# TODO: hacky but works ;D maybe rename latents somehow? # TODO: hacky but works ;D maybe rename latents somehow?
context.services.latents.set(name_positive, (c, c_ec)) context.services.latents.set(name_positive, (c, ec))
context.services.latents.set(name_negative, (uc, uc_ec)) context.services.latents.set(name_negative, (uc, None))
return CompelOutput( return CompelOutput(
positive=ConditioningField( positive=ConditioningField(

2
invokeai/backend/model_management/model_manager.py
View File

@ -1191,6 +1191,8 @@ class ModelManager(object):
if self.embedding_path is not None: if self.embedding_path is not None:
print(f">> Loading embeddings from {self.embedding_path}") print(f">> Loading embeddings from {self.embedding_path}")
for root, _, files in os.walk(self.embedding_path): for root, _, files in os.walk(self.embedding_path):
print(root)
print(files)
for name in files: for name in files:
ti_path = os.path.join(root, name) ti_path = os.path.join(root, name)
model.textual_inversion_manager.load_textual_inversion( model.textual_inversion_manager.load_textual_inversion(