stable-diffusion-webui/modules/sd_models_xl.py

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

109 lines
4.6 KiB
Python
Raw Normal View History

2023-07-11 12:16:43 -06:00
from __future__ import annotations
import torch
import sgm.models.diffusion
import sgm.modules.diffusionmodules.denoiser_scaling
import sgm.modules.diffusionmodules.discretizer
2023-07-12 14:52:43 -06:00
from modules import devices, shared, prompt_parser
2023-07-11 12:16:43 -06:00
2023-07-12 14:52:43 -06:00
def get_learned_conditioning(self: sgm.models.diffusion.DiffusionEngine, batch: prompt_parser.SdConditioning | list[str]):
2023-07-11 12:16:43 -06:00
for embedder in self.conditioner.embedders:
embedder.ucg_rate = 0.0
width = getattr(batch, 'width', 1024)
height = getattr(batch, 'height', 1024)
2023-07-14 00:16:01 -06:00
is_negative_prompt = getattr(batch, 'is_negative_prompt', False)
aesthetic_score = shared.opts.sdxl_refiner_low_aesthetic_score if is_negative_prompt else shared.opts.sdxl_refiner_high_aesthetic_score
devices_args = dict(device=devices.device, dtype=devices.dtype)
2023-07-12 14:52:43 -06:00
sdxl_conds = {
"txt": batch,
2023-07-14 00:16:01 -06:00
"original_size_as_tuple": torch.tensor([height, width], **devices_args).repeat(len(batch), 1),
"crop_coords_top_left": torch.tensor([shared.opts.sdxl_crop_top, shared.opts.sdxl_crop_left], **devices_args).repeat(len(batch), 1),
"target_size_as_tuple": torch.tensor([height, width], **devices_args).repeat(len(batch), 1),
"aesthetic_score": torch.tensor([aesthetic_score], **devices_args).repeat(len(batch), 1),
2023-07-12 14:52:43 -06:00
}
2023-07-14 00:16:01 -06:00
force_zero_negative_prompt = is_negative_prompt and all(x == '' for x in batch)
c = self.conditioner(sdxl_conds, force_zero_embeddings=['txt'] if force_zero_negative_prompt else [])
2023-07-11 12:16:43 -06:00
return c
def apply_model(self: sgm.models.diffusion.DiffusionEngine, x, t, cond):
return self.model(x, t, cond)
2023-07-13 07:18:39 -06:00
def get_first_stage_encoding(self, x): # SDXL's encode_first_stage does everything so get_first_stage_encoding is just there for compatibility
return x
2023-07-14 00:16:01 -06:00
sgm.models.diffusion.DiffusionEngine.get_learned_conditioning = get_learned_conditioning
sgm.models.diffusion.DiffusionEngine.apply_model = apply_model
sgm.models.diffusion.DiffusionEngine.get_first_stage_encoding = get_first_stage_encoding
def encode_embedding_init_text(self: sgm.modules.GeneralConditioner, init_text, nvpt):
res = []
for embedder in [embedder for embedder in self.embedders if hasattr(embedder, 'encode_embedding_init_text')]:
encoded = embedder.encode_embedding_init_text(init_text, nvpt)
res.append(encoded)
return torch.cat(res, dim=1)
2023-07-29 06:15:06 -06:00
def tokenize(self: sgm.modules.GeneralConditioner, texts):
for embedder in [embedder for embedder in self.embedders if hasattr(embedder, 'tokenize')]:
return embedder.tokenize(texts)
raise AssertionError('no tokenizer available')
2023-07-14 00:16:01 -06:00
def process_texts(self, texts):
for embedder in [embedder for embedder in self.embedders if hasattr(embedder, 'process_texts')]:
return embedder.process_texts(texts)
def get_target_prompt_token_count(self, token_count):
for embedder in [embedder for embedder in self.embedders if hasattr(embedder, 'get_target_prompt_token_count')]:
return embedder.get_target_prompt_token_count(token_count)
# those additions to GeneralConditioner make it possible to use it as model.cond_stage_model from SD1.5 in exist
sgm.modules.GeneralConditioner.encode_embedding_init_text = encode_embedding_init_text
2023-07-29 06:15:06 -06:00
sgm.modules.GeneralConditioner.tokenize = tokenize
2023-07-14 00:16:01 -06:00
sgm.modules.GeneralConditioner.process_texts = process_texts
sgm.modules.GeneralConditioner.get_target_prompt_token_count = get_target_prompt_token_count
2023-07-11 12:16:43 -06:00
def extend_sdxl(model):
2023-07-14 00:16:01 -06:00
"""this adds a bunch of parameters to make SDXL model look a bit more like SD1.5 to the rest of the codebase."""
2023-07-11 12:16:43 -06:00
dtype = next(model.model.diffusion_model.parameters()).dtype
model.model.diffusion_model.dtype = dtype
model.model.conditioning_key = 'crossattn'
2023-07-14 00:16:01 -06:00
model.cond_stage_key = 'txt'
# model.cond_stage_model will be set in sd_hijack
2023-07-11 12:16:43 -06:00
model.parameterization = "v" if isinstance(model.denoiser.scaling, sgm.modules.diffusionmodules.denoiser_scaling.VScaling) else "eps"
discretization = sgm.modules.diffusionmodules.discretizer.LegacyDDPMDiscretization()
model.alphas_cumprod = torch.asarray(discretization.alphas_cumprod, device=devices.device, dtype=dtype)
2023-07-14 00:16:01 -06:00
model.conditioner.wrapped = torch.nn.Module()
2023-07-12 14:52:43 -06:00
2023-07-11 12:16:43 -06:00
sgm.modules.attention.print = shared.ldm_print
sgm.modules.diffusionmodules.model.print = shared.ldm_print
sgm.modules.diffusionmodules.openaimodel.print = shared.ldm_print
sgm.modules.encoders.modules.print = shared.ldm_print
2023-07-12 14:52:43 -06:00
2023-07-13 00:30:33 -06:00
# this gets the code to load the vanilla attention that we override
sgm.modules.attention.SDP_IS_AVAILABLE = True
2023-07-13 00:38:54 -06:00
sgm.modules.attention.XFORMERS_IS_AVAILABLE = False