SDXL support

Author: AUTOMATIC1111

modules/launch_utils.py

@@ -224,6 +224,20 @@ def run_extensions_installers(settings_file):
         run_extension_installer(os.path.join(extensions_dir, dirname_extension))
 
 
+def mute_sdxl_imports():
+    """create fake modules that SDXL wants to import but doesn't actually use for our purposes"""
+
+    import importlib
+
+    module = importlib.util.module_from_spec(importlib.machinery.ModuleSpec('taming.modules.losses.lpips', None))
+    module.LPIPS = None
+    sys.modules['taming.modules.losses.lpips'] = module
+
+    module = importlib.util.module_from_spec(importlib.machinery.ModuleSpec('sgm.data', None))
+    module.StableDataModuleFromConfig = None
+    sys.modules['sgm.data'] = module
+
+
 def prepare_environment():
     torch_index_url = os.environ.get('TORCH_INDEX_URL', "https://download.pytorch.org/whl/cu118")
     torch_command = os.environ.get('TORCH_COMMAND', f"pip install torch==2.0.1 torchvision==0.15.2 --extra-index-url {torch_index_url}")
@@ -319,11 +333,14 @@ def prepare_environment():
     if args.update_all_extensions:
         git_pull_recursive(extensions_dir)
 
+    mute_sdxl_imports()
+
     if "--exit" in sys.argv:
         print("Exiting because of --exit argument")
         exit(0)
 
 
+
 def configure_for_tests():
     if "--api" not in sys.argv:
         sys.argv.append("--api")

modules/lowvram.py

@@ -53,19 +53,46 @@ def first_stage_model_decode_wrap(z):
         send_me_to_gpu(first_stage_model, None)
         return first_stage_model_decode(z)
 
-    # for SD1, cond_stage_model is CLIP and its NN is in the tranformer frield, but for SD2, it's open clip, and it's in model field
-    if hasattr(sd_model.cond_stage_model, 'model'):
-        sd_model.cond_stage_model.transformer = sd_model.cond_stage_model.model
-
-    # remove several big modules: cond, first_stage, depth/embedder (if applicable), and unet from the model and then
-    # send the model to GPU. Then put modules back. the modules will be in CPU.
-    stored = sd_model.cond_stage_model.transformer, sd_model.first_stage_model, getattr(sd_model, 'depth_model', None), getattr(sd_model, 'embedder', None), sd_model.model
-    sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.depth_model, sd_model.embedder, sd_model.model = None, None, None, None, None
+    to_remain_in_cpu = [
+        (sd_model, 'first_stage_model'),
+        (sd_model, 'depth_model'),
+        (sd_model, 'embedder'),
+        (sd_model, 'model'),
+        (sd_model, 'embedder'),
+    ]
+
+    is_sdxl = hasattr(sd_model, 'conditioner')
+    is_sd2 = not is_sdxl and hasattr(sd_model.cond_stage_model, 'model')
+
+    if is_sdxl:
+        to_remain_in_cpu.append((sd_model, 'conditioner'))
+    elif is_sd2:
+        to_remain_in_cpu.append((sd_model.cond_stage_model, 'model'))
+    else:
+        to_remain_in_cpu.append((sd_model.cond_stage_model, 'transformer'))
+
+    # remove several big modules: cond, first_stage, depth/embedder (if applicable), and unet from the model
+    stored = []
+    for obj, field in to_remain_in_cpu:
+        module = getattr(obj, field, None)
+        stored.append(module)
+        setattr(obj, field, None)
+
+    # send the model to GPU.
     sd_model.to(devices.device)
-    sd_model.cond_stage_model.transformer, sd_model.first_stage_model, sd_model.depth_model, sd_model.embedder, sd_model.model = stored
+
+    # put modules back. the modules will be in CPU.
+    for (obj, field), module in zip(to_remain_in_cpu, stored):
+        setattr(obj, field, module)
 
     # register hooks for those the first three models
-    sd_model.cond_stage_model.transformer.register_forward_pre_hook(send_me_to_gpu)
+    if is_sdxl:
+        sd_model.conditioner.register_forward_pre_hook(send_me_to_gpu)
+    elif is_sd2:
+        sd_model.cond_stage_model.model.register_forward_pre_hook(send_me_to_gpu)
+    else:
+        sd_model.cond_stage_model.transformer.register_forward_pre_hook(send_me_to_gpu)
+
     sd_model.first_stage_model.register_forward_pre_hook(send_me_to_gpu)
     sd_model.first_stage_model.encode = first_stage_model_encode_wrap
     sd_model.first_stage_model.decode = first_stage_model_decode_wrap
@@ -75,10 +102,6 @@ def first_stage_model_decode_wrap(z):
         sd_model.embedder.register_forward_pre_hook(send_me_to_gpu)
     parents[sd_model.cond_stage_model.transformer] = sd_model.cond_stage_model
 
-    if hasattr(sd_model.cond_stage_model, 'model'):
-        sd_model.cond_stage_model.model = sd_model.cond_stage_model.transformer
-        del sd_model.cond_stage_model.transformer
-
     if use_medvram:
         sd_model.model.register_forward_pre_hook(send_me_to_gpu)
     else:

modules/paths.py

@@ -20,7 +20,7 @@
 
 path_dirs = [
     (sd_path, 'ldm', 'Stable Diffusion', []),
-    (os.path.join(sd_path, '../generative-models'), 'sgm', 'Stable Diffusion XL', []),
+    (os.path.join(sd_path, '../generative-models'), 'sgm', 'Stable Diffusion XL', ["sgm"]),
     (os.path.join(sd_path, '../CodeFormer'), 'inference_codeformer.py', 'CodeFormer', []),
     (os.path.join(sd_path, '../BLIP'), 'models/blip.py', 'BLIP', []),
     (os.path.join(sd_path, '../k-diffusion'), 'k_diffusion/sampling.py', 'k_diffusion', ["atstart"]),
@@ -36,6 +36,13 @@
         d = os.path.abspath(d)
         if "atstart" in options:
             sys.path.insert(0, d)
+        elif "sgm" in options:
+            # Stable Diffusion XL repo has scripts dir with __init__.py in it which ruins every extension's scripts dir, so we
+            # import sgm and remove it from sys.path so that when a script imports scripts.something, it doesbn't use sgm's scripts dir.
+
+            sys.path.insert(0, d)
+            import sgm
+            sys.path.pop(0)
         else:
             sys.path.append(d)
         paths[what] = d

modules/processing.py

@@ -343,10 +343,13 @@ def get_conds_with_caching(self, function, required_prompts, steps, caches, extr
         return cache[1]
 
     def setup_conds(self):
+        prompts = prompt_parser.SdConditioning(self.prompts, width=self.width, height=self.height)
+        negative_prompts = prompt_parser.SdConditioning(self.negative_prompts, width=self.width, height=self.height)
+
         sampler_config = sd_samplers.find_sampler_config(self.sampler_name)
         self.step_multiplier = 2 if sampler_config and sampler_config.options.get("second_order", False) else 1
-        self.uc = self.get_conds_with_caching(prompt_parser.get_learned_conditioning, self.negative_prompts, self.steps * self.step_multiplier, [self.cached_uc], self.extra_network_data)
-        self.c = self.get_conds_with_caching(prompt_parser.get_multicond_learned_conditioning, self.prompts, self.steps * self.step_multiplier, [self.cached_c], self.extra_network_data)
+        self.uc = self.get_conds_with_caching(prompt_parser.get_learned_conditioning, negative_prompts, self.steps * self.step_multiplier, [self.cached_uc], self.extra_network_data)
+        self.c = self.get_conds_with_caching(prompt_parser.get_multicond_learned_conditioning, prompts, self.steps * self.step_multiplier, [self.cached_c], self.extra_network_data)
 
     def parse_extra_network_prompts(self):
         self.prompts, self.extra_network_data = extra_networks.parse_prompts(self.prompts)

modules/prompt_parser.py

@@ -1,3 +1,5 @@
+from __future__ import annotations
+
 import re
 from collections import namedtuple
 from typing import List
@@ -109,7 +111,19 @@ def get_schedule(prompt):
 ScheduledPromptConditioning = namedtuple("ScheduledPromptConditioning", ["end_at_step", "cond"])
 
 
-def get_learned_conditioning(model, prompts, steps):
+class SdConditioning(list):
+    """
+    A list with prompts for stable diffusion's conditioner model.
+    Can also specify width and height of created image - SDXL needs it.
+    """
+    def __init__(self, prompts, width=None, height=None):
+        super().__init__()
+        self.extend(prompts)
+        self.width = width or getattr(prompts, 'width', None)
+        self.height = height or getattr(prompts, 'height', None)
+
+
+def get_learned_conditioning(model, prompts: SdConditioning | list[str], steps):
     """converts a list of prompts into a list of prompt schedules - each schedule is a list of ScheduledPromptConditioning, specifying the comdition (cond),
     and the sampling step at which this condition is to be replaced by the next one.
 
@@ -160,11 +174,13 @@ def get_learned_conditioning(model, prompts, steps):
 re_AND = re.compile(r"\bAND\b")
 re_weight = re.compile(r"^(.*?)(?:\s*:\s*([-+]?(?:\d+\.?|\d*\.\d+)))?\s*$")
 
-def get_multicond_prompt_list(prompts):
+
+def get_multicond_prompt_list(prompts: SdConditioning | list[str]):
     res_indexes = []
 
-    prompt_flat_list = []
     prompt_indexes = {}
+    prompt_flat_list = SdConditioning(prompts)
+    prompt_flat_list.clear()
 
     for prompt in prompts:
         subprompts = re_AND.split(prompt)
@@ -201,6 +217,7 @@ def __init__(self, shape, batch):
         self.shape: tuple = shape  # the shape field is needed to send this object to DDIM/PLMS
         self.batch: List[List[ComposableScheduledPromptConditioning]] = batch
 
+
 def get_multicond_learned_conditioning(model, prompts, steps) -> MulticondLearnedConditioning:
     """same as get_learned_conditioning, but returns a list of ScheduledPromptConditioning along with the weight objects for each prompt.
     For each prompt, the list is obtained by splitting the prompt using the AND separator.

modules/sd_hijack.py

@@ -15,6 +15,11 @@
 import ldm.models.diffusion.plms
 import ldm.modules.encoders.modules
 
+import sgm.modules.attention
+import sgm.modules.diffusionmodules.model
+import sgm.modules.diffusionmodules.openaimodel
+import sgm.modules.encoders.modules
+
 attention_CrossAttention_forward = ldm.modules.attention.CrossAttention.forward
 diffusionmodules_model_nonlinearity = ldm.modules.diffusionmodules.model.nonlinearity
 diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.AttnBlock.forward
@@ -56,6 +61,9 @@ def apply_optimizations(option=None):
     ldm.modules.diffusionmodules.model.nonlinearity = silu
     ldm.modules.diffusionmodules.openaimodel.th = sd_hijack_unet.th
 
+    sgm.modules.diffusionmodules.model.nonlinearity = silu
+    sgm.modules.diffusionmodules.openaimodel.th = sd_hijack_unet.th
+
     if current_optimizer is not None:
         current_optimizer.undo()
         current_optimizer = None
@@ -89,6 +97,10 @@ def undo_optimizations():
     ldm.modules.attention.CrossAttention.forward = hypernetwork.attention_CrossAttention_forward
     ldm.modules.diffusionmodules.model.AttnBlock.forward = diffusionmodules_model_AttnBlock_forward
 
+    sgm.modules.diffusionmodules.model.nonlinearity = diffusionmodules_model_nonlinearity
+    sgm.modules.attention.CrossAttention.forward = hypernetwork.attention_CrossAttention_forward
+    sgm.modules.diffusionmodules.model.AttnBlock.forward = diffusionmodules_model_AttnBlock_forward
+
 
 def fix_checkpoint():
     """checkpoints are now added and removed in embedding/hypernet code, since torch doesn't want
@@ -170,10 +182,19 @@ def hijack(self, m):
         if conditioner:
             for i in range(len(conditioner.embedders)):
                 embedder = conditioner.embedders[i]
-                if type(embedder).__name__ == 'FrozenOpenCLIPEmbedder':
+                typename = type(embedder).__name__
+                if typename == 'FrozenOpenCLIPEmbedder':
                     embedder.model.token_embedding = EmbeddingsWithFixes(embedder.model.token_embedding, self)
                     m.cond_stage_model = sd_hijack_open_clip.FrozenOpenCLIPEmbedderWithCustomWords(embedder, self)
                     conditioner.embedders[i] = m.cond_stage_model
+                if typename == 'FrozenCLIPEmbedder':
+                    model_embeddings = m.cond_stage_model.transformer.text_model.embeddings
+                    model_embeddings.token_embedding = EmbeddingsWithFixes(model_embeddings.token_embedding, self)
+                    m.cond_stage_model = sd_hijack_clip.FrozenCLIPEmbedderWithCustomWords(embedder, self)
+                    conditioner.embedders[i] = m.cond_stage_model
+                if typename == 'FrozenOpenCLIPEmbedder2':
+                    embedder.model.token_embedding = EmbeddingsWithFixes(embedder.model.token_embedding, self)
+                    conditioner.embedders[i] = sd_hijack_open_clip.FrozenOpenCLIPEmbedder2WithCustomWords(embedder, self)
 
         if type(m.cond_stage_model) == xlmr.BertSeriesModelWithTransformation:
             model_embeddings = m.cond_stage_model.roberta.embeddings

modules/sd_hijack_clip.py

@@ -42,6 +42,10 @@ def __init__(self, wrapped, hijack):
         self.hijack: sd_hijack.StableDiffusionModelHijack = hijack
         self.chunk_length = 75
 
+        self.is_trainable = getattr(wrapped, 'is_trainable', False)
+        self.input_key = getattr(wrapped, 'input_key', 'txt')
+        self.legacy_ucg_val = None
+
     def empty_chunk(self):
         """creates an empty PromptChunk and returns it"""
 
@@ -199,8 +203,9 @@ def forward(self, texts):
         """
         Accepts an array of texts; Passes texts through transformers network to create a tensor with numerical representation of those texts.
         Returns a tensor with shape of (B, T, C), where B is length of the array; T is length, in tokens, of texts (including padding) - T will
-        be a multiple of 77; and C is dimensionality of each token - for SD1 it's 768, and for SD2 it's 1024.
+        be a multiple of 77; and C is dimensionality of each token - for SD1 it's 768, for SD2 it's 1024, and for SDXL it's 1280.
         An example shape returned by this function can be: (2, 77, 768).
+        For SDXL, instead of returning one tensor avobe, it returns a tuple with two: the other one with shape (B, 1280) with pooled values.
         Webui usually sends just one text at a time through this function - the only time when texts is an array with more than one elemenet
         is when you do prompt editing: "a picture of a [cat:dog:0.4] eating ice cream"
         """
@@ -233,7 +238,10 @@ def forward(self, texts):
             embeddings_list = ", ".join([f'{name} [{embedding.checksum()}]' for name, embedding in used_embeddings.items()])
             self.hijack.comments.append(f"Used embeddings: {embeddings_list}")
 
-        return torch.hstack(zs)
+        if getattr(self.wrapped, 'return_pooled', False):
+            return torch.hstack(zs), zs[0].pooled
+        else:
+            return torch.hstack(zs)
 
     def process_tokens(self, remade_batch_tokens, batch_multipliers):
         """
@@ -256,9 +264,9 @@ def process_tokens(self, remade_batch_tokens, batch_multipliers):
         # restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
         batch_multipliers = torch.asarray(batch_multipliers).to(devices.device)
         original_mean = z.mean()
-        z = z * batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
+        z *= batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
         new_mean = z.mean()
-        z = z * (original_mean / new_mean)
+        z *= (original_mean / new_mean)
 
         return z
 

modules/sd_hijack_open_clip.py

@@ -16,10 +16,6 @@ def __init__(self, wrapped, hijack):
         self.id_end = tokenizer.encoder["<end_of_text>"]
         self.id_pad = 0
 
-        self.is_trainable = getattr(wrapped, 'is_trainable', False)
-        self.input_key = getattr(wrapped, 'input_key', 'txt')
-        self.legacy_ucg_val = None
-
     def tokenize(self, texts):
         assert not opts.use_old_emphasis_implementation, 'Old emphasis implementation not supported for Open Clip'
 
@@ -39,3 +35,37 @@ def encode_embedding_init_text(self, init_text, nvpt):
         embedded = self.wrapped.model.token_embedding.wrapped(ids).squeeze(0)
 
         return embedded
+
+
+class FrozenOpenCLIPEmbedder2WithCustomWords(sd_hijack_clip.FrozenCLIPEmbedderWithCustomWordsBase):
+    def __init__(self, wrapped, hijack):
+        super().__init__(wrapped, hijack)
+
+        self.comma_token = [v for k, v in tokenizer.encoder.items() if k == ',</w>'][0]
+        self.id_start = tokenizer.encoder["<start_of_text>"]
+        self.id_end = tokenizer.encoder["<end_of_text>"]
+        self.id_pad = 0
+
+    def tokenize(self, texts):
+        assert not opts.use_old_emphasis_implementation, 'Old emphasis implementation not supported for Open Clip'
+
+        tokenized = [tokenizer.encode(text) for text in texts]
+
+        return tokenized
+
+    def encode_with_transformers(self, tokens):
+        d = self.wrapped.encode_with_transformer(tokens)
+        z = d[self.wrapped.layer]
+
+        pooled = d.get("pooled")
+        if pooled is not None:
+            z.pooled = pooled
+
+        return z
+
+    def encode_embedding_init_text(self, init_text, nvpt):
+        ids = tokenizer.encode(init_text)
+        ids = torch.asarray([ids], device=devices.device, dtype=torch.int)
+        embedded = self.wrapped.model.token_embedding.wrapped(ids).squeeze(0)
+
+        return embedded