init

README.md

@@ -70,6 +70,7 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web
 - No token limit for prompts (original stable diffusion lets you use up to 75 tokens)
 - DeepDanbooru integration, creates danbooru style tags for anime prompts (add --deepdanbooru to commandline args)
 - [xformers](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Xformers), major speed increase for select cards: (add --xformers to commandline args)
+- Aesthetic, a way to generate images with a specific aesthetic by using clip images embds (implementation of https://github.com/vicgalle/stable-diffusion-aesthetic-gradients)
 
 ## Installation and Running
 Make sure the required [dependencies](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Dependencies) are met and follow the instructions available for both [NVidia](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-NVidia-GPUs) (recommended) and [AMD](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs) GPUs.

modules/processing.py

@@ -316,9 +316,14 @@ def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments, iteration
     return f"{all_prompts[index]}{negative_prompt_text}\n{generation_params_text}".strip()
 
 
-def process_images(p: StableDiffusionProcessing) -> Processed:
+def process_images(p: StableDiffusionProcessing, aesthetic_lr=0, aesthetic_weight=0, aesthetic_steps=0,
+                   aesthetic_imgs=None,aesthetic_slerp=False) -> Processed:
     """this is the main loop that both txt2img and img2img use; it calls func_init once inside all the scopes and func_sample once per batch"""
 
+    aesthetic_lr = float(aesthetic_lr)
+    aesthetic_weight = float(aesthetic_weight)
+    aesthetic_steps = int(aesthetic_steps)
+
     if type(p.prompt) == list:
         assert (len(p.prompt) > 0)
     else:
@@ -394,7 +399,13 @@ def process_images(p: StableDiffusionProcessing) -> Processed:
             #uc = p.sd_model.get_learned_conditioning(len(prompts) * [p.negative_prompt])
             #c = p.sd_model.get_learned_conditioning(prompts)
             with devices.autocast():
-                uc = prompt_parser.get_learned_conditioning(shared.sd_model, len(prompts) * [p.negative_prompt], p.steps)
+                if hasattr(shared.sd_model.cond_stage_model, "set_aesthetic_params"):
+                    shared.sd_model.cond_stage_model.set_aesthetic_params(0, 0, 0)
+                uc = prompt_parser.get_learned_conditioning(shared.sd_model, len(prompts) * [p.negative_prompt],
+                                                            p.steps)
+                if hasattr(shared.sd_model.cond_stage_model, "set_aesthetic_params"):
+                    shared.sd_model.cond_stage_model.set_aesthetic_params(aesthetic_lr, aesthetic_weight,
+                                                                          aesthetic_steps, aesthetic_imgs,aesthetic_slerp)
                 c = prompt_parser.get_multicond_learned_conditioning(shared.sd_model, prompts, p.steps)
 
             if len(model_hijack.comments) > 0:

modules/sd_hijack.py

@@ -9,11 +9,14 @@ from torch.nn.functional import silu
 
 import modules.textual_inversion.textual_inversion
 from modules import prompt_parser, devices, sd_hijack_optimizations, shared
-from modules.shared import opts, device, cmd_opts
+from modules.shared import opts, device, cmd_opts, aesthetic_embeddings
 from modules.sd_hijack_optimizations import invokeAI_mps_available
 
 import ldm.modules.attention
 import ldm.modules.diffusionmodules.model
+from transformers import CLIPVisionModel, CLIPModel
+import torch.optim as optim
+import copy
 
 attention_CrossAttention_forward = ldm.modules.attention.CrossAttention.forward
 diffusionmodules_model_nonlinearity = ldm.modules.diffusionmodules.model.nonlinearity
@@ -109,13 +112,29 @@ class StableDiffusionModelHijack:
         _, remade_batch_tokens, _, _, _, token_count = self.clip.process_text([text])
         return remade_batch_tokens[0], token_count, get_target_prompt_token_count(token_count)
 
+def slerp(low, high, val):
+    low_norm = low/torch.norm(low, dim=1, keepdim=True)
+    high_norm = high/torch.norm(high, dim=1, keepdim=True)
+    omega = torch.acos((low_norm*high_norm).sum(1))
+    so = torch.sin(omega)
+    res = (torch.sin((1.0-val)*omega)/so).unsqueeze(1)*low + (torch.sin(val*omega)/so).unsqueeze(1) * high
+    return res
 
 class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
     def __init__(self, wrapped, hijack):
         super().__init__()
         self.wrapped = wrapped
+        self.clipModel = CLIPModel.from_pretrained(
+            self.wrapped.transformer.name_or_path
+        )
+        del self.clipModel.vision_model
         self.hijack: StableDiffusionModelHijack = hijack
         self.tokenizer = wrapped.tokenizer
+        # self.vision = CLIPVisionModel.from_pretrained(self.wrapped.transformer.name_or_path).eval()
+        self.image_embs_name = None
+        self.image_embs = None
+        self.load_image_embs(None)
+
         self.token_mults = {}
 
         self.comma_token = [v for k, v in self.tokenizer.get_vocab().items() if k == ',</w>'][0]
@@ -136,6 +155,23 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
             if mult != 1.0:
                 self.token_mults[ident] = mult
 
+    def set_aesthetic_params(self, aesthetic_lr, aesthetic_weight, aesthetic_steps, image_embs_name=None,
+                             aesthetic_slerp=True):
+        self.slerp = aesthetic_slerp
+        self.aesthetic_lr = aesthetic_lr
+        self.aesthetic_weight = aesthetic_weight
+        self.aesthetic_steps = aesthetic_steps
+        self.load_image_embs(image_embs_name)
+
+    def load_image_embs(self, image_embs_name):
+        if image_embs_name is None or len(image_embs_name) == 0:
+            image_embs_name = None
+        if image_embs_name is not None and self.image_embs_name != image_embs_name:
+            self.image_embs_name = image_embs_name
+            self.image_embs = torch.load(aesthetic_embeddings[self.image_embs_name], map_location=device)
+            self.image_embs /= self.image_embs.norm(dim=-1, keepdim=True)
+            self.image_embs.requires_grad_(False)
+
     def tokenize_line(self, line, used_custom_terms, hijack_comments):
         id_end = self.wrapped.tokenizer.eos_token_id
 
@@ -334,6 +370,46 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
             z1 = self.process_tokens(tokens, multipliers)
             z = z1 if z is None else torch.cat((z, z1), axis=-2)
 
+            if len(text[
+                       0]) != 0 and self.aesthetic_steps != 0 and self.aesthetic_lr != 0 and self.aesthetic_weight != 0 and self.image_embs_name != None:
+                if not opts.use_old_emphasis_implementation:
+                    remade_batch_tokens = [
+                        [self.wrapped.tokenizer.bos_token_id] + x[:75] + [self.wrapped.tokenizer.eos_token_id] for x in
+                        remade_batch_tokens]
+
+                tokens = torch.asarray(remade_batch_tokens).to(device)
+                with torch.enable_grad():
+                    model = copy.deepcopy(self.clipModel).to(device)
+                    model.requires_grad_(True)
+
+                    # We optimize the model to maximize the similarity
+                    optimizer = optim.Adam(
+                        model.text_model.parameters(), lr=self.aesthetic_lr
+                    )
+
+                    for i in range(self.aesthetic_steps):
+                        text_embs = model.get_text_features(input_ids=tokens)
+                        text_embs = text_embs / text_embs.norm(dim=-1, keepdim=True)
+                        sim = text_embs @ self.image_embs.T
+                        loss = -sim
+                        optimizer.zero_grad()
+                        loss.mean().backward()
+                        optimizer.step()
+
+                    zn = model.text_model(input_ids=tokens, output_hidden_states=-opts.CLIP_stop_at_last_layers)
+                    if opts.CLIP_stop_at_last_layers > 1:
+                        zn = zn.hidden_states[-opts.CLIP_stop_at_last_layers]
+                        zn = model.text_model.final_layer_norm(zn)
+                    else:
+                        zn = zn.last_hidden_state
+                    model.cpu()
+                    del model
+
+                if self.slerp:
+                    z = slerp(z, zn, self.aesthetic_weight)
+                else:
+                    z = z * (1 - self.aesthetic_weight) + zn * self.aesthetic_weight
+
             remade_batch_tokens = rem_tokens
             batch_multipliers = rem_multipliers
             i += 1

modules/shared.py

@@ -30,6 +30,8 @@ parser.add_argument("--no-half-vae", action='store_true', help="do not switch th
 parser.add_argument("--no-progressbar-hiding", action='store_true', help="do not hide progressbar in gradio UI (we hide it because it slows down ML if you have hardware acceleration in browser)")
 parser.add_argument("--max-batch-count", type=int, default=16, help="maximum batch count value for the UI")
 parser.add_argument("--embeddings-dir", type=str, default=os.path.join(script_path, 'embeddings'), help="embeddings directory for textual inversion (default: embeddings)")
+parser.add_argument("--aesthetic_embeddings-dir", type=str, default=os.path.join(script_path, 'aesthetic_embeddings'),
+                    help="aesthetic_embeddings directory(default: aesthetic_embeddings)")
 parser.add_argument("--hypernetwork-dir", type=str, default=os.path.join(models_path, 'hypernetworks'), help="hypernetwork directory")
 parser.add_argument("--allow-code", action='store_true', help="allow custom script execution from webui")
 parser.add_argument("--medvram", action='store_true', help="enable stable diffusion model optimizations for sacrificing a little speed for low VRM usage")
@@ -90,6 +92,9 @@ os.makedirs(cmd_opts.hypernetwork_dir, exist_ok=True)
 hypernetworks = hypernetwork.list_hypernetworks(cmd_opts.hypernetwork_dir)
 loaded_hypernetwork = None
 
+aesthetic_embeddings = {f.replace(".pt",""): os.path.join(cmd_opts.aesthetic_embeddings_dir, f) for f in
+                        os.listdir(cmd_opts.aesthetic_embeddings_dir) if f.endswith(".pt")}
+
 
 def reload_hypernetworks():
     global hypernetworks

modules/textual_inversion/dataset.py

@@ -48,7 +48,7 @@ class PersonalizedBase(Dataset):
         print("Preparing dataset...")
         for path in tqdm.tqdm(self.image_paths):
             try:
-                image = Image.open(path).convert('RGB').resize((self.width, self.height), PIL.Image.BICUBIC)
+                image = Image.open(path).convert('RGB').resize((self.width, self.height), PIL.Image.Resampling.BICUBIC)
             except Exception:
                 continue
 

modules/textual_inversion/textual_inversion.py

@@ -172,7 +172,15 @@ def create_embedding(name, num_vectors_per_token, init_text='*'):
     return fn
 
 
-def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps, create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding, preview_image_prompt):
+def batched(dataset, total, n=1):
+    for ndx in range(0, total, n):
+        yield [dataset.__getitem__(i) for i in range(ndx, min(ndx + n, total))]
+
+
+def train_embedding(embedding_name, learn_rate, data_root, log_directory, training_width, training_height, steps,
+                    create_image_every, save_embedding_every, template_file, save_image_with_stored_embedding,
+                    preview_image_prompt, batch_size=1,
+                    gradient_accumulation=1):
     assert embedding_name, 'embedding not selected'
 
     shared.state.textinfo = "Initializing textual inversion training..."
@@ -204,7 +212,11 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini
 
     shared.state.textinfo = f"Preparing dataset from {html.escape(data_root)}..."
     with torch.autocast("cuda"):
-        ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width, height=training_height, repeats=shared.opts.training_image_repeats_per_epoch, placeholder_token=embedding_name, model=shared.sd_model, device=devices.device, template_file=template_file)
+        ds = modules.textual_inversion.dataset.PersonalizedBase(data_root=data_root, width=training_width,
+                                                                height=training_height,
+                                                                repeats=shared.opts.training_image_repeats_per_epoch,
+                                                                placeholder_token=embedding_name, model=shared.sd_model,
+                                                                device=devices.device, template_file=template_file)
 
     hijack = sd_hijack.model_hijack
 
@@ -223,7 +235,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini
     scheduler = LearnRateScheduler(learn_rate, steps, ititial_step)
     optimizer = torch.optim.AdamW([embedding.vec], lr=scheduler.learn_rate)
 
-    pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step)
+    pbar = tqdm.tqdm(enumerate(batched(ds, steps - ititial_step, batch_size)), total=steps - ititial_step)
     for i, entry in pbar:
         embedding.step = i + ititial_step
 
@@ -235,17 +247,20 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini
             break
 
         with torch.autocast("cuda"):
-            c = cond_model([entry.cond_text])
+            c = cond_model([e.cond_text for e in entry])
+
+            x = torch.stack([e.latent for e in entry]).to(devices.device)
+            loss = shared.sd_model(x, c)[0]
 
-            x = entry.latent.to(devices.device)
-            loss = shared.sd_model(x.unsqueeze(0), c)[0]
             del x
 
             losses[embedding.step % losses.shape[0]] = loss.item()
 
-            optimizer.zero_grad()
             loss.backward()
+            if ((i + 1) % gradient_accumulation == 0) or (i + 1 == steps - ititial_step):
                 optimizer.step()
+                optimizer.zero_grad()
+
 
         epoch_num = embedding.step // len(ds)
         epoch_step = embedding.step - (epoch_num * len(ds)) + 1
@@ -259,7 +274,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini
         if embedding.step > 0 and images_dir is not None and embedding.step % create_image_every == 0:
             last_saved_image = os.path.join(images_dir, f'{embedding_name}-{embedding.step}.png')
 
-            preview_text = entry.cond_text if preview_image_prompt == "" else preview_image_prompt
+            preview_text = entry[0].cond_text if preview_image_prompt == "" else preview_image_prompt
 
             p = processing.StableDiffusionProcessingTxt2Img(
                 sd_model=shared.sd_model,
@@ -305,7 +320,7 @@ def train_embedding(embedding_name, learn_rate, data_root, log_directory, traini
 <p>
 Loss: {losses.mean():.7f}<br/>
 Step: {embedding.step}<br/>
-Last prompt: {html.escape(entry.cond_text)}<br/>
+Last prompt: {html.escape(entry[-1].cond_text)}<br/>
 Last saved embedding: {html.escape(last_saved_file)}<br/>
 Last saved image: {html.escape(last_saved_image)}<br/>
 </p>

modules/txt2img.py

@@ -6,7 +6,14 @@ import modules.processing as processing
 from modules.ui import plaintext_to_html
 
 
-def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: str, steps: int, sampler_index: int, restore_faces: bool, tiling: bool, n_iter: int, batch_size: int, cfg_scale: float, seed: int, subseed: int, subseed_strength: float, seed_resize_from_h: int, seed_resize_from_w: int, seed_enable_extras: bool, height: int, width: int, enable_hr: bool, scale_latent: bool, denoising_strength: float, *args):
+def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2: str, steps: int, sampler_index: int,
+            restore_faces: bool, tiling: bool, n_iter: int, batch_size: int, cfg_scale: float, seed: int, subseed: int,
+            subseed_strength: float, seed_resize_from_h: int, seed_resize_from_w: int, seed_enable_extras: bool,
+            height: int, width: int, enable_hr: bool, scale_latent: bool, denoising_strength: float,
+            aesthetic_lr=0,
+            aesthetic_weight=0, aesthetic_steps=0,
+            aesthetic_imgs=None,
+            aesthetic_slerp=False, *args):
     p = StableDiffusionProcessingTxt2Img(
         sd_model=shared.sd_model,
         outpath_samples=opts.outdir_samples or opts.outdir_txt2img_samples,
@@ -40,7 +47,7 @@ def txt2img(prompt: str, negative_prompt: str, prompt_style: str, prompt_style2:
     processed = modules.scripts.scripts_txt2img.run(p, *args)
 
     if processed is None:
-        processed = process_images(p)
+        processed = process_images(p, aesthetic_lr, aesthetic_weight, aesthetic_steps, aesthetic_imgs, aesthetic_slerp)
 
     shared.total_tqdm.clear()
 

modules/ui.py

@@ -24,7 +24,8 @@ import gradio.routes
 
 from modules import sd_hijack
 from modules.paths import script_path
-from modules.shared import opts, cmd_opts
+from modules.shared import opts, cmd_opts,aesthetic_embeddings
+
 if cmd_opts.deepdanbooru:
     from modules.deepbooru import get_deepbooru_tags
 import modules.shared as shared
@@ -534,6 +535,14 @@ def create_ui(wrap_gradio_gpu_call):
                     width = gr.Slider(minimum=64, maximum=2048, step=64, label="Width", value=512)
                     height = gr.Slider(minimum=64, maximum=2048, step=64, label="Height", value=512)
 
+                with gr.Group():
+                    aesthetic_lr = gr.Textbox(label='Learning rate', placeholder="Learning rate", value="0.005")
+                    aesthetic_weight = gr.Slider(minimum=0, maximum=1, step=0.01, label="Aesthetic weight", value=0.7)
+                    aesthetic_steps = gr.Slider(minimum=0, maximum=50, step=1, label="Aesthetic steps", value=50)
+
+                    aesthetic_imgs = gr.Dropdown(sorted(aesthetic_embeddings.keys()), label="Imgs embedding", value=sorted(aesthetic_embeddings.keys())[0] if len(aesthetic_embeddings) > 0 else None)
+                    aesthetic_slerp = gr.Checkbox(label="Slerp interpolation", value=False)
+
                 with gr.Row():
                     restore_faces = gr.Checkbox(label='Restore faces', value=False, visible=len(shared.face_restorers) > 1)
                     tiling = gr.Checkbox(label='Tiling', value=False)
@@ -604,6 +613,11 @@ def create_ui(wrap_gradio_gpu_call):
                            enable_hr,
                            scale_latent,
                            denoising_strength,
+                           aesthetic_lr,
+                           aesthetic_weight,
+                           aesthetic_steps,
+                           aesthetic_imgs,
+                           aesthetic_slerp
                        ] + custom_inputs,
                 outputs=[
                     txt2img_gallery,
@@ -1097,6 +1111,9 @@ def create_ui(wrap_gradio_gpu_call):
                     template_file = gr.Textbox(label='Prompt template file', value=os.path.join(script_path, "textual_inversion_templates", "style_filewords.txt"))
                     training_width = gr.Slider(minimum=64, maximum=2048, step=64, label="Width", value=512)
                     training_height = gr.Slider(minimum=64, maximum=2048, step=64, label="Height", value=512)
+                    batch_size = gr.Slider(minimum=1, maximum=64, step=1, label="Batch Size", value=4)
+                    gradient_accumulation = gr.Slider(minimum=1, maximum=256, step=1, label="Gradient accumulation",
+                                                      value=1)
                     steps = gr.Number(label='Max steps', value=100000, precision=0)
                     create_image_every = gr.Number(label='Save an image to log directory every N steps, 0 to disable', value=500, precision=0)
                     save_embedding_every = gr.Number(label='Save a copy of embedding to log directory every N steps, 0 to disable', value=500, precision=0)
@@ -1180,6 +1197,8 @@ def create_ui(wrap_gradio_gpu_call):
                 template_file,
                 save_image_with_stored_embedding,
                 preview_image_prompt,
+                batch_size,
+                gradient_accumulation
             ],
             outputs=[
                 ti_output,