hypernetwork training mk1

So `[job_timestamp]` pattern can use in saving image UI.

README.md

@@ -16,6 +16,7 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web
 - Attention, specify parts of text that the model should pay more attention to
     - a man in a ((tuxedo)) - will pay more attention to tuxedo
     - a man in a (tuxedo:1.21) - alternative syntax
+    - select text and press ctrl+up or ctrl+down to aduotmatically adjust attention to selected text
 - Loopback, run img2img processing multiple times
 - X/Y plot, a way to draw a 2 dimensional plot of images with different parameters
 - Textual Inversion
@@ -61,6 +62,9 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web
 - Reloading checkpoints on the fly
 - Checkpoint Merger, a tab that allows you to merge two checkpoints into one
 - [Custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Custom-Scripts) with many extensions from community
+- [Composable-Diffusion](https://energy-based-model.github.io/Compositional-Visual-Generation-with-Composable-Diffusion-Models/), a way to use multiple prompts at once
+     - separate prompts using uppercase `AND`
+     - also supports weights for prompts: `a cat :1.2 AND a dog AND a penguin :2.2`
 
 ## 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.

javascript/edit-attention.js

@@ -0,0 +1,41 @@
+addEventListener('keydown', (event) => {
+	let target = event.originalTarget;
+	if (!target.hasAttribute("placeholder")) return;
+	if (!target.placeholder.toLowerCase().includes("prompt")) return;
+
+	let plus = "ArrowUp"
+	let minus = "ArrowDown"
+	if (event.key != plus && event.key != minus) return;
+
+	selectionStart = target.selectionStart;
+	selectionEnd = target.selectionEnd;
+	if(selectionStart == selectionEnd) return;
+
+	event.preventDefault();
+
+	if (selectionStart == 0 || target.value[selectionStart - 1] != "(") {
+		target.value = target.value.slice(0, selectionStart) +
+			"(" + target.value.slice(selectionStart, selectionEnd) + ":1.0)" +
+			target.value.slice(selectionEnd);
+
+		target.focus();
+		target.selectionStart = selectionStart + 1;
+		target.selectionEnd = selectionEnd + 1;
+
+	} else {
+		end = target.value.slice(selectionEnd + 1).indexOf(")") + 1;
+		weight = parseFloat(target.value.slice(selectionEnd + 1, selectionEnd + 1 + end));
+		if (event.key == minus) weight -= 0.1;
+		if (event.key == plus) weight += 0.1;
+
+		weight = parseFloat(weight.toPrecision(12));
+
+		target.value = target.value.slice(0, selectionEnd + 1) +
+			weight +
+			target.value.slice(selectionEnd + 1 + end - 1);
+
+		target.focus();
+		target.selectionStart = selectionStart;
+		target.selectionEnd = selectionEnd;
+	}
+});

javascript/progressbar.js

@@ -5,7 +5,7 @@ function check_progressbar(id_part, id_progressbar, id_progressbar_span, id_inte
     var progressbar = gradioApp().getElementById(id_progressbar)
     var interrupt = gradioApp().getElementById(id_interrupt)
     
-    if(progressbar && progressbar.offsetParent){
+    if(opts.show_progress_in_title && progressbar && progressbar.offsetParent){
         if(progressbar.innerText){
             let newtitle = 'Stable Diffusion - ' + progressbar.innerText
             if(document.title != newtitle){

launch.py

@@ -89,7 +89,7 @@ def git_clone(url, dir, name, commithash=None):
         if commithash is None:
             return
 
-        current_hash = run(f'"{git}" -C {dir} rev-parse HEAD', None, "Couldn't determine {name}'s hash: {commithash}").strip()
+        current_hash = run(f'"{git}" -C {dir} rev-parse HEAD', None, f"Couldn't determine {name}'s hash: {commithash}").strip()
         if current_hash == commithash:
             return
 

modules/extras.py

@@ -100,6 +100,8 @@ def run_extras(extras_mode, image, image_folder, gfpgan_visibility, codeformer_v
 
         outputs.append(image)
 
+    devices.torch_gc()
+
     return outputs, plaintext_to_html(info), ''
 
 

modules/hypernetwork/hypernetwork.py

@@ -0,0 +1,267 @@
+import datetime
+import glob
+import html
+import os
+import sys
+import traceback
+import tqdm
+
+import torch
+
+from ldm.util import default
+from modules import devices, shared, processing, sd_models
+import torch
+from torch import einsum
+from einops import rearrange, repeat
+import modules.textual_inversion.dataset
+
+
+class HypernetworkModule(torch.nn.Module):
+    def __init__(self, dim, state_dict=None):
+        super().__init__()
+
+        self.linear1 = torch.nn.Linear(dim, dim * 2)
+        self.linear2 = torch.nn.Linear(dim * 2, dim)
+
+        if state_dict is not None:
+            self.load_state_dict(state_dict, strict=True)
+        else:
+            self.linear1.weight.data.fill_(0.0001)
+            self.linear1.bias.data.fill_(0.0001)
+            self.linear2.weight.data.fill_(0.0001)
+            self.linear2.bias.data.fill_(0.0001)
+
+        self.to(devices.device)
+
+    def forward(self, x):
+        return x + (self.linear2(self.linear1(x)))
+
+
+class Hypernetwork:
+    filename = None
+    name = None
+
+    def __init__(self, name=None):
+        self.filename = None
+        self.name = name
+        self.layers = {}
+        self.step = 0
+        self.sd_checkpoint = None
+        self.sd_checkpoint_name = None
+
+        for size in [320, 640, 768, 1280]:
+            self.layers[size] = (HypernetworkModule(size), HypernetworkModule(size))
+
+    def weights(self):
+        res = []
+
+        for k, layers in self.layers.items():
+            for layer in layers:
+                layer.train()
+                res += [layer.linear1.weight, layer.linear1.bias, layer.linear2.weight, layer.linear2.bias]
+
+        return res
+
+    def save(self, filename):
+        state_dict = {}
+
+        for k, v in self.layers.items():
+            state_dict[k] = (v[0].state_dict(), v[1].state_dict())
+
+        state_dict['step'] = self.step
+        state_dict['name'] = self.name
+        state_dict['sd_checkpoint'] = self.sd_checkpoint
+        state_dict['sd_checkpoint_name'] = self.sd_checkpoint_name
+
+        torch.save(state_dict, filename)
+
+    def load(self, filename):
+        self.filename = filename
+        if self.name is None:
+            self.name = os.path.splitext(os.path.basename(filename))[0]
+
+        state_dict = torch.load(filename, map_location='cpu')
+
+        for size, sd in state_dict.items():
+            if type(size) == int:
+                self.layers[size] = (HypernetworkModule(size, sd[0]), HypernetworkModule(size, sd[1]))
+
+        self.name = state_dict.get('name', self.name)
+        self.step = state_dict.get('step', 0)
+        self.sd_checkpoint = state_dict.get('sd_checkpoint', None)
+        self.sd_checkpoint_name = state_dict.get('sd_checkpoint_name', None)
+
+
+def load_hypernetworks(path):
+    res = {}
+
+    for filename in glob.iglob(path + '**/*.pt', recursive=True):
+        try:
+            hn = Hypernetwork()
+            hn.load(filename)
+            res[hn.name] = hn
+        except Exception:
+            print(f"Error loading hypernetwork {filename}", file=sys.stderr)
+            print(traceback.format_exc(), file=sys.stderr)
+
+    return res
+
+
+def attention_CrossAttention_forward(self, x, context=None, mask=None):
+    h = self.heads
+
+    q = self.to_q(x)
+    context = default(context, x)
+
+    hypernetwork_layers = (shared.hypernetwork.layers if shared.hypernetwork is not None else {}).get(context.shape[2], None)
+
+    if hypernetwork_layers is not None:
+        hypernetwork_k, hypernetwork_v = hypernetwork_layers
+
+        self.hypernetwork_k = hypernetwork_k
+        self.hypernetwork_v = hypernetwork_v
+
+        context_k = hypernetwork_k(context)
+        context_v = hypernetwork_v(context)
+    else:
+        context_k = context
+        context_v = context
+
+    k = self.to_k(context_k)
+    v = self.to_v(context_v)
+
+    q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
+
+    sim = einsum('b i d, b j d -> b i j', q, k) * self.scale
+
+    if mask is not None:
+        mask = rearrange(mask, 'b ... -> b (...)')
+        max_neg_value = -torch.finfo(sim.dtype).max
+        mask = repeat(mask, 'b j -> (b h) () j', h=h)
+        sim.masked_fill_(~mask, max_neg_value)
+
+    # attention, what we cannot get enough of
+    attn = sim.softmax(dim=-1)
+
+    out = einsum('b i j, b j d -> b i d', attn, v)
+    out = rearrange(out, '(b h) n d -> b n (h d)', h=h)
+    return self.to_out(out)
+
+
+def train_hypernetwork(hypernetwork_name, learn_rate, data_root, log_directory, steps, create_image_every, save_hypernetwork_every, template_file, preview_image_prompt):
+    assert hypernetwork_name, 'embedding not selected'
+
+    shared.hypernetwork = shared.hypernetworks[hypernetwork_name]
+
+    shared.state.textinfo = "Initializing hypernetwork training..."
+    shared.state.job_count = steps
+
+    filename = os.path.join(shared.cmd_opts.hypernetwork_dir, f'{hypernetwork_name}.pt')
+
+    log_directory = os.path.join(log_directory, datetime.datetime.now().strftime("%Y-%m-%d"), hypernetwork_name)
+
+    if save_hypernetwork_every > 0:
+        hypernetwork_dir = os.path.join(log_directory, "hypernetworks")
+        os.makedirs(hypernetwork_dir, exist_ok=True)
+    else:
+        hypernetwork_dir = None
+
+    if create_image_every > 0:
+        images_dir = os.path.join(log_directory, "images")
+        os.makedirs(images_dir, exist_ok=True)
+    else:
+        images_dir = None
+
+    cond_model = shared.sd_model.cond_stage_model
+
+    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, size=512, placeholder_token=hypernetwork_name, model=shared.sd_model, device=devices.device, template_file=template_file)
+
+    hypernetwork = shared.hypernetworks[hypernetwork_name]
+    weights = hypernetwork.weights()
+    for weight in weights:
+        weight.requires_grad = True
+
+    optimizer = torch.optim.AdamW(weights, lr=learn_rate)
+
+    losses = torch.zeros((32,))
+
+    last_saved_file = "<none>"
+    last_saved_image = "<none>"
+
+    ititial_step = hypernetwork.step or 0
+    if ititial_step > steps:
+        return hypernetwork, filename
+
+    pbar = tqdm.tqdm(enumerate(ds), total=steps-ititial_step)
+    for i, (x, text) in pbar:
+        hypernetwork.step = i + ititial_step
+
+        if hypernetwork.step > steps:
+            break
+
+        if shared.state.interrupted:
+            break
+
+        with torch.autocast("cuda"):
+            c = cond_model([text])
+
+            x = x.to(devices.device)
+            loss = shared.sd_model(x.unsqueeze(0), c)[0]
+            del x
+
+            losses[hypernetwork.step % losses.shape[0]] = loss.item()
+
+            optimizer.zero_grad()
+            loss.backward()
+            optimizer.step()
+
+        pbar.set_description(f"loss: {losses.mean():.7f}")
+
+        if hypernetwork.step > 0 and hypernetwork_dir is not None and hypernetwork.step % save_hypernetwork_every == 0:
+            last_saved_file = os.path.join(hypernetwork_dir, f'{hypernetwork_name}-{hypernetwork.step}.pt')
+            hypernetwork.save(last_saved_file)
+
+        if hypernetwork.step > 0 and images_dir is not None and hypernetwork.step % create_image_every == 0:
+            last_saved_image = os.path.join(images_dir, f'{hypernetwork_name}-{hypernetwork.step}.png')
+
+            preview_text = text if preview_image_prompt == "" else preview_image_prompt
+
+            p = processing.StableDiffusionProcessingTxt2Img(
+                sd_model=shared.sd_model,
+                prompt=preview_text,
+                steps=20,
+                do_not_save_grid=True,
+                do_not_save_samples=True,
+            )
+
+            processed = processing.process_images(p)
+            image = processed.images[0]
+
+            shared.state.current_image = image
+            image.save(last_saved_image)
+
+            last_saved_image += f", prompt: {preview_text}"
+
+        shared.state.job_no = hypernetwork.step
+
+        shared.state.textinfo = f"""
+<p>
+Loss: {losses.mean():.7f}<br/>
+Step: {hypernetwork.step}<br/>
+Last prompt: {html.escape(text)}<br/>
+Last saved embedding: {html.escape(last_saved_file)}<br/>
+Last saved image: {html.escape(last_saved_image)}<br/>
+</p>
+"""
+
+    checkpoint = sd_models.select_checkpoint()
+
+    hypernetwork.sd_checkpoint = checkpoint.hash
+    hypernetwork.sd_checkpoint_name = checkpoint.model_name
+    hypernetwork.save(filename)
+
+    return hypernetwork, filename
+
+

modules/hypernetwork/ui.py

@@ -0,0 +1,43 @@
+import html
+import os
+
+import gradio as gr
+
+import modules.textual_inversion.textual_inversion
+import modules.textual_inversion.preprocess
+from modules import sd_hijack, shared
+
+
+def create_hypernetwork(name):
+    fn = os.path.join(shared.cmd_opts.hypernetwork_dir, f"{name}.pt")
+    assert not os.path.exists(fn), f"file {fn} already exists"
+
+    hypernetwork = modules.hypernetwork.hypernetwork.Hypernetwork(name=name)
+    hypernetwork.save(fn)
+
+    shared.reload_hypernetworks()
+    shared.hypernetwork = shared.hypernetworks.get(shared.opts.sd_hypernetwork, None)
+
+    return gr.Dropdown.update(choices=sorted([x for x in shared.hypernetworks.keys()])), f"Created: {fn}", ""
+
+
+def train_hypernetwork(*args):
+
+    initial_hypernetwork = shared.hypernetwork
+
+    try:
+        sd_hijack.undo_optimizations()
+
+        hypernetwork, filename = modules.hypernetwork.hypernetwork.train_hypernetwork(*args)
+
+        res = f"""
+Training {'interrupted' if shared.state.interrupted else 'finished'} at {hypernetwork.step} steps.
+Hypernetwork saved to {html.escape(filename)}
+"""
+        return res, ""
+    except Exception:
+        raise
+    finally:
+        shared.hypernetwork = initial_hypernetwork
+        sd_hijack.apply_optimizations()
+

modules/images.py

@@ -292,18 +292,13 @@ def apply_filename_pattern(x, p, seed, prompt):
         x = x.replace("[cfg]", str(p.cfg_scale))
         x = x.replace("[width]", str(p.width))
         x = x.replace("[height]", str(p.height))
-
-        #currently disabled if using the save button, will work otherwise
-        # if enabled it will cause a bug because styles is not included in the save_files data dictionary
-        if hasattr(p, "styles"):
-            x = x.replace("[styles]", sanitize_filename_part(", ".join([x for x in p.styles if not x == "None"]) or "None", replace_spaces=False))
-
+        x = x.replace("[styles]", sanitize_filename_part(", ".join([x for x in p.styles if not x == "None"]) or "None", replace_spaces=False))
         x = x.replace("[sampler]", sanitize_filename_part(sd_samplers.samplers[p.sampler_index].name, replace_spaces=False))
 
-    x = x.replace("[model_hash]", shared.sd_model.sd_model_hash)
+    x = x.replace("[model_hash]", getattr(p, "sd_model_hash", shared.sd_model.sd_model_hash))
     x = x.replace("[date]", datetime.date.today().isoformat())
     x = x.replace("[datetime]", datetime.datetime.now().strftime("%Y%m%d%H%M%S"))
-    x = x.replace("[job_timestamp]", shared.state.job_timestamp)
+    x = x.replace("[job_timestamp]", getattr(p, "job_timestamp", shared.state.job_timestamp))
 
     # Apply [prompt] at last. Because it may contain any replacement word.^M
     if prompt is not None:
@@ -353,7 +348,7 @@ def get_next_sequence_number(path, basename):
     return result + 1
 
 
-def save_image(image, path, basename, seed=None, prompt=None, extension='png', info=None, short_filename=False, no_prompt=False, grid=False, pnginfo_section_name='parameters', p=None, existing_info=None, forced_filename=None, suffix=""):
+def save_image(image, path, basename, seed=None, prompt=None, extension='png', info=None, short_filename=False, no_prompt=False, grid=False, pnginfo_section_name='parameters', p=None, existing_info=None, forced_filename=None, suffix="", save_to_dirs=None):
     if short_filename or prompt is None or seed is None:
         file_decoration = ""
     elif opts.save_to_dirs:
@@ -377,7 +372,8 @@ def save_image(image, path, basename, seed=None, prompt=None, extension='png', i
     else:
         pnginfo = None
 
-    save_to_dirs = (grid and opts.grid_save_to_dirs) or (not grid and opts.save_to_dirs and not no_prompt)
+    if save_to_dirs is None:
+        save_to_dirs = (grid and opts.grid_save_to_dirs) or (not grid and opts.save_to_dirs and not no_prompt)
 
     if save_to_dirs:
         dirname = apply_filename_pattern(opts.directories_filename_pattern or "[prompt_words]", p, seed, prompt).strip('\\ /')
@@ -431,4 +427,4 @@ def save_image(image, path, basename, seed=None, prompt=None, extension='png', i
         with open(f"{fullfn_without_extension}.txt", "w", encoding="utf8") as file:
             file.write(info + "\n")
 
-
+    return fullfn

modules/processing.py

@@ -11,7 +11,7 @@ import cv2
 from skimage import exposure
 
 import modules.sd_hijack
-from modules import devices, prompt_parser, masking, sd_samplers
+from modules import devices, prompt_parser, masking, sd_samplers, lowvram
 from modules.sd_hijack import model_hijack
 from modules.shared import opts, cmd_opts, state
 import modules.shared as shared
@@ -121,6 +121,8 @@ class Processed:
         self.denoising_strength = getattr(p, 'denoising_strength', None)
         self.extra_generation_params = p.extra_generation_params
         self.index_of_first_image = index_of_first_image
+        self.styles = p.styles
+        self.job_timestamp = state.job_timestamp
 
         self.eta = p.eta
         self.ddim_discretize = p.ddim_discretize
@@ -165,6 +167,8 @@ class Processed:
             "extra_generation_params": self.extra_generation_params,
             "index_of_first_image": self.index_of_first_image,
             "infotexts": self.infotexts,
+            "styles": self.styles,
+            "job_timestamp": self.job_timestamp,
         }
 
         return json.dumps(obj)
@@ -382,6 +386,13 @@ def process_images(p: StableDiffusionProcessing) -> Processed:
             x_samples_ddim = p.sd_model.decode_first_stage(samples_ddim)
             x_samples_ddim = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)
 
+            del samples_ddim
+
+            if shared.cmd_opts.lowvram or shared.cmd_opts.medvram:
+                lowvram.send_everything_to_cpu()
+
+            devices.torch_gc()
+
             if opts.filter_nsfw:
                 import modules.safety as safety
                 x_samples_ddim = modules.safety.censor_batch(x_samples_ddim)
@@ -423,9 +434,15 @@ def process_images(p: StableDiffusionProcessing) -> Processed:
                 if opts.samples_save and not p.do_not_save_samples:
                     images.save_image(image, p.outpath_samples, "", seeds[i], prompts[i], opts.samples_format, info=infotext(n, i), p=p)
 
-                infotexts.append(infotext(n, i))
+                text = infotext(n, i)
+                infotexts.append(text)
+                image.info["parameters"] = text
                 output_images.append(image)
 
+            del x_samples_ddim 
+
+            devices.torch_gc()
+
             state.nextjob()
 
         p.color_corrections = None
@@ -436,7 +453,9 @@ def process_images(p: StableDiffusionProcessing) -> Processed:
             grid = images.image_grid(output_images, p.batch_size)
 
             if opts.return_grid:
-                infotexts.insert(0, infotext())
+                text = infotext()
+                infotexts.insert(0, text)
+                grid.info["parameters"] = text
                 output_images.insert(0, grid)
                 index_of_first_image = 1
 
@@ -663,4 +682,7 @@ class StableDiffusionProcessingImg2Img(StableDiffusionProcessing):
         if self.mask is not None:
             samples = samples * self.nmask + self.init_latent * self.mask
 
+        del x
+        devices.torch_gc()
+
         return samples

modules/sd_hijack.py

@@ -5,6 +5,7 @@ import traceback
 import torch
 import numpy as np
 from torch import einsum
+from torch.nn.functional import silu
 
 import modules.textual_inversion.textual_inversion
 from modules import prompt_parser, devices, sd_hijack_optimizations, shared
@@ -19,16 +20,21 @@ diffusionmodules_model_AttnBlock_forward = ldm.modules.diffusionmodules.model.At
 
 
 def apply_optimizations():
+    undo_optimizations()
+
+    ldm.modules.diffusionmodules.model.nonlinearity = silu
+
     if cmd_opts.opt_split_attention_v1:
         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_v1
     elif not cmd_opts.disable_opt_split_attention and (cmd_opts.opt_split_attention or torch.cuda.is_available()):
         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward
-        ldm.modules.diffusionmodules.model.nonlinearity = sd_hijack_optimizations.nonlinearity_hijack
         ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.cross_attention_attnblock_forward
 
 
 def undo_optimizations():
-    ldm.modules.attention.CrossAttention.forward = attention_CrossAttention_forward
+    from modules.hypernetwork import hypernetwork
+
+    ldm.modules.attention.CrossAttention.forward = hypernetwork.attention_CrossAttention_forward
     ldm.modules.diffusionmodules.model.nonlinearity = diffusionmodules_model_nonlinearity
     ldm.modules.diffusionmodules.model.AttnBlock.forward = diffusionmodules_model_AttnBlock_forward
 

modules/sd_hijack_optimizations.py

@@ -5,6 +5,8 @@ from torch import einsum
 from ldm.util import default
 from einops import rearrange
 
+from modules import shared
+
 
 # see https://github.com/basujindal/stable-diffusion/pull/117 for discussion
 def split_cross_attention_forward_v1(self, x, context=None, mask=None):
@@ -42,8 +44,18 @@ def split_cross_attention_forward(self, x, context=None, mask=None):
 
     q_in = self.to_q(x)
     context = default(context, x)
-    k_in = self.to_k(context) * self.scale
-    v_in = self.to_v(context)
+
+    hypernetwork_layers = (shared.hypernetwork.layers if shared.hypernetwork is not None else {}).get(context.shape[2], None)
+
+    if hypernetwork_layers is not None:
+        k_in = self.to_k(hypernetwork_layers[0](context))
+        v_in = self.to_v(hypernetwork_layers[1](context))
+    else:
+        k_in = self.to_k(context)
+        v_in = self.to_v(context)
+
+    k_in *= self.scale
+
     del context, x
 
     q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q_in, k_in, v_in))
@@ -92,14 +104,6 @@ def split_cross_attention_forward(self, x, context=None, mask=None):
 
     return self.to_out(r2)
 
-def nonlinearity_hijack(x):
-    # swish
-    t = torch.sigmoid(x)
-    x *= t
-    del t
-
-    return x
-
 def cross_attention_attnblock_forward(self, x):
         h_ = x
         h_ = self.norm(h_)

modules/sd_models.py

@@ -134,6 +134,14 @@ def load_model_weights(model, checkpoint_file, sd_model_hash):
 
     devices.dtype = torch.float32 if shared.cmd_opts.no_half else torch.float16
 
+    vae_file = os.path.splitext(checkpoint_file)[0] + ".vae.pt"
+    if os.path.exists(vae_file):
+        print(f"Loading VAE weights from: {vae_file}")
+        vae_ckpt = torch.load(vae_file, map_location="cpu")
+        vae_dict = {k: v for k, v in vae_ckpt["state_dict"].items() if k[0:4] != "loss"}
+
+        model.first_stage_model.load_state_dict(vae_dict)
+
     model.sd_model_hash = sd_model_hash
     model.sd_model_checkpint = checkpoint_file
 

modules/sd_samplers.py

@@ -338,9 +338,11 @@ class KDiffusionSampler:
         steps, t_enc = setup_img2img_steps(p, steps)
 
         if p.sampler_noise_scheduler_override:
-          sigmas = p.sampler_noise_scheduler_override(steps)
+            sigmas = p.sampler_noise_scheduler_override(steps)
+        elif self.config is not None and self.config.options.get('scheduler', None) == 'karras':
+            sigmas = k_diffusion.sampling.get_sigmas_karras(n=steps, sigma_min=0.1, sigma_max=10, device=shared.device)
         else:
-          sigmas = self.model_wrap.get_sigmas(steps)
+            sigmas = self.model_wrap.get_sigmas(steps)
 
         noise = noise * sigmas[steps - t_enc - 1]
         xi = x + noise

modules/shared.py

@@ -14,11 +14,10 @@ import modules.sd_models
 import modules.styles
 import modules.devices as devices
 from modules import sd_samplers
-from modules.paths import script_path, sd_path
+from modules.paths import models_path, script_path, sd_path
 
 sd_model_file = os.path.join(script_path, 'model.ckpt')
 default_sd_model_file = sd_model_file
-model_path = os.path.join(script_path, 'models')
 parser = argparse.ArgumentParser()
 parser.add_argument("--config", type=str, default=os.path.join(sd_path, "configs/stable-diffusion/v1-inference.yaml"), help="path to config which constructs model",)
 parser.add_argument("--ckpt", type=str, default=sd_model_file, help="path to checkpoint of stable diffusion model; if specified, this checkpoint will be added to the list of checkpoints and loaded",)
@@ -29,6 +28,7 @@ parser.add_argument("--no-half", action='store_true', help="do not switch the mo
 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("--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")
 parser.add_argument("--lowvram", action='store_true', help="enable stable diffusion model optimizations for sacrificing a lot of speed for very low VRM usage")
@@ -36,14 +36,14 @@ parser.add_argument("--always-batch-cond-uncond", action='store_true', help="dis
 parser.add_argument("--unload-gfpgan", action='store_true', help="does not do anything.")
 parser.add_argument("--precision", type=str, help="evaluate at this precision", choices=["full", "autocast"], default="autocast")
 parser.add_argument("--share", action='store_true', help="use share=True for gradio and make the UI accessible through their site (doesn't work for me but you might have better luck)")
-parser.add_argument("--codeformer-models-path", type=str, help="Path to directory with codeformer model file(s).", default=os.path.join(model_path, 'Codeformer'))
-parser.add_argument("--gfpgan-models-path", type=str, help="Path to directory with GFPGAN model file(s).", default=os.path.join(model_path, 'GFPGAN'))
-parser.add_argument("--esrgan-models-path", type=str, help="Path to directory with ESRGAN model file(s).", default=os.path.join(model_path, 'ESRGAN'))
-parser.add_argument("--bsrgan-models-path", type=str, help="Path to directory with BSRGAN model file(s).", default=os.path.join(model_path, 'BSRGAN'))
-parser.add_argument("--realesrgan-models-path", type=str, help="Path to directory with RealESRGAN model file(s).", default=os.path.join(model_path, 'RealESRGAN'))
-parser.add_argument("--scunet-models-path", type=str, help="Path to directory with ScuNET model file(s).", default=os.path.join(model_path, 'ScuNET'))
-parser.add_argument("--swinir-models-path", type=str, help="Path to directory with SwinIR model file(s).", default=os.path.join(model_path, 'SwinIR'))
-parser.add_argument("--ldsr-models-path", type=str, help="Path to directory with LDSR model file(s).", default=os.path.join(model_path, 'LDSR'))
+parser.add_argument("--codeformer-models-path", type=str, help="Path to directory with codeformer model file(s).", default=os.path.join(models_path, 'Codeformer'))
+parser.add_argument("--gfpgan-models-path", type=str, help="Path to directory with GFPGAN model file(s).", default=os.path.join(models_path, 'GFPGAN'))
+parser.add_argument("--esrgan-models-path", type=str, help="Path to directory with ESRGAN model file(s).", default=os.path.join(models_path, 'ESRGAN'))
+parser.add_argument("--bsrgan-models-path", type=str, help="Path to directory with BSRGAN model file(s).", default=os.path.join(models_path, 'BSRGAN'))
+parser.add_argument("--realesrgan-models-path", type=str, help="Path to directory with RealESRGAN model file(s).", default=os.path.join(models_path, 'RealESRGAN'))
+parser.add_argument("--scunet-models-path", type=str, help="Path to directory with ScuNET model file(s).", default=os.path.join(models_path, 'ScuNET'))
+parser.add_argument("--swinir-models-path", type=str, help="Path to directory with SwinIR model file(s).", default=os.path.join(models_path, 'SwinIR'))
+parser.add_argument("--ldsr-models-path", type=str, help="Path to directory with LDSR model file(s).", default=os.path.join(models_path, 'LDSR'))
 parser.add_argument("--opt-split-attention", action='store_true', help="force-enables cross-attention layer optimization. By default, it's on for torch.cuda and off for other torch devices.")
 parser.add_argument("--disable-opt-split-attention", action='store_true', help="force-disables cross-attention layer optimization")
 parser.add_argument("--opt-split-attention-v1", action='store_true', help="enable older version of split attention optimization that does not consume all the VRAM it can find")
@@ -78,6 +78,16 @@ parallel_processing_allowed = not cmd_opts.lowvram and not cmd_opts.medvram
 config_filename = cmd_opts.ui_settings_file
 
 
+def reload_hypernetworks():
+    from modules.hypernetwork import hypernetwork
+    hypernetworks.clear()
+    hypernetworks.update(hypernetwork.load_hypernetworks(cmd_opts.hypernetwork_dir))
+
+
+hypernetworks = {}
+hypernetwork = None
+
+
 class State:
     interrupted = False
     job = ""
@@ -207,6 +217,7 @@ options_templates.update(options_section(('system', "System"), {
 
 options_templates.update(options_section(('sd', "Stable Diffusion"), {
     "sd_model_checkpoint": OptionInfo(None, "Stable Diffusion checkpoint", gr.Dropdown, lambda: {"choices": modules.sd_models.checkpoint_tiles()}),
+    "sd_hypernetwork": OptionInfo("None", "Stable Diffusion finetune hypernetwork", gr.Dropdown, lambda: {"choices": ["None"] + [x for x in hypernetworks.keys()]}),
     "img2img_color_correction": OptionInfo(False, "Apply color correction to img2img results to match original colors."),
     "save_images_before_color_correction": OptionInfo(False, "Save a copy of image before applying color correction to img2img results"),
     "img2img_fix_steps": OptionInfo(False, "With img2img, do exactly the amount of steps the slider specifies (normally you'd do less with less denoising)."),
@@ -236,6 +247,7 @@ options_templates.update(options_section(('ui', "User interface"), {
     "font": OptionInfo("", "Font for image grids that have text"),
     "js_modal_lightbox": OptionInfo(True, "Enable full page image viewer"),
     "js_modal_lightbox_initialy_zoomed": OptionInfo(True, "Show images zoomed in by default in full page image viewer"),
+    "show_progress_in_title": OptionInfo(True, "Show generation progress in window title."),
 }))
 
 options_templates.update(options_section(('sampler-params', "Sampler parameters"), {

modules/textual_inversion/ui.py

@@ -22,7 +22,6 @@ def preprocess(*args):
 
 
 def train_embedding(*args):
-
     try:
         sd_hijack.undo_optimizations()
 

modules/ui.py

@@ -35,8 +35,9 @@ import modules.codeformer_model
 import modules.styles
 import modules.generation_parameters_copypaste
 from modules import prompt_parser
-from modules.images import apply_filename_pattern, get_next_sequence_number
+from modules.images import save_image
 import modules.textual_inversion.ui
+import modules.hypernetwork.ui
 
 # this is a fix for Windows users. Without it, javascript files will be served with text/html content-type and the bowser will not show any UI
 mimetypes.init()
@@ -114,20 +115,13 @@ def save_files(js_data, images, index):
     p = MyObject(data)
     path = opts.outdir_save
     save_to_dirs = opts.use_save_to_dirs_for_ui
-
-    if save_to_dirs:
-        dirname = apply_filename_pattern(opts.directories_filename_pattern or "[prompt_words]", p, p.seed, p.prompt)
-        path = os.path.join(opts.outdir_save, dirname)
-
-    os.makedirs(path, exist_ok=True)
-  
+    extension: str = opts.samples_format
+    start_index = 0
 
     if index > -1 and opts.save_selected_only and (index >= data["index_of_first_image"]):  # ensures we are looking at a specific non-grid picture, and we have save_selected_only
 
         images = [images[index]]
-        infotexts = [data["infotexts"][index]]
-    else:
-        infotexts = data["infotexts"]
+        start_index = index
 
     with open(os.path.join(opts.outdir_save, "log.csv"), "a", encoding="utf8", newline='') as file:
         at_start = file.tell() == 0
@@ -135,37 +129,18 @@ def save_files(js_data, images, index):
         if at_start:
             writer.writerow(["prompt", "seed", "width", "height", "sampler", "cfgs", "steps", "filename", "negative_prompt"])
 
-        file_decoration = opts.samples_filename_pattern or "[seed]-[prompt_spaces]"
-        if file_decoration != "":
-            file_decoration = "-" + file_decoration.lower()
-        file_decoration = apply_filename_pattern(file_decoration, p, p.seed, p.prompt)
-        truncated = (file_decoration[:240] + '..') if len(file_decoration) > 240 else file_decoration
-        filename_base = truncated
-        extension = opts.samples_format.lower()
-
-        basecount = get_next_sequence_number(path, "")
-        for i, filedata in enumerate(images):
-            file_number = f"{basecount+i:05}"
-            filename = file_number + filename_base + f".{extension}"
-            filepath = os.path.join(path, filename)
-
-
+        for image_index, filedata in enumerate(images, start_index):
             if filedata.startswith("data:image/png;base64,"):
                 filedata = filedata[len("data:image/png;base64,"):]
 
             image = Image.open(io.BytesIO(base64.decodebytes(filedata.encode('utf-8'))))
-            if opts.enable_pnginfo and extension == 'png':
-                pnginfo = PngImagePlugin.PngInfo()
-                pnginfo.add_text('parameters', infotexts[i])
-                image.save(filepath, pnginfo=pnginfo)
-            else:
-                image.save(filepath, quality=opts.jpeg_quality)
 
-            if opts.enable_pnginfo and extension in ("jpg", "jpeg", "webp"):
-                piexif.insert(piexif.dump({"Exif": {
-                    piexif.ExifIFD.UserComment: piexif.helper.UserComment.dump(infotexts[i], encoding="unicode")
-                }}), filepath)
+            is_grid = image_index < p.index_of_first_image
+            i = 0 if is_grid else (image_index - p.index_of_first_image)
 
+            fullfn = save_image(image, path, "", seed=p.all_seeds[i], prompt=p.all_prompts[i], extension=extension, info=p.infotexts[image_index], grid=is_grid, p=p, save_to_dirs=save_to_dirs)
+
+            filename = os.path.relpath(fullfn, path)
             filenames.append(filename)
 
         writer.writerow([data["prompt"], data["seed"], data["width"], data["height"], data["sampler"], data["cfg_scale"], data["steps"], filenames[0], data["negative_prompt"]])
@@ -991,6 +966,18 @@ def create_ui(wrap_gradio_gpu_call):
                         with gr.Column():
                             create_embedding = gr.Button(value="Create", variant='primary')
 
+                with gr.Group():
+                    gr.HTML(value="<p style='margin-bottom: 0.7em'>Create a new hypernetwork</p>")
+
+                    new_hypernetwork_name = gr.Textbox(label="Name")
+
+                    with gr.Row():
+                        with gr.Column(scale=3):
+                            gr.HTML(value="")
+
+                        with gr.Column():
+                            create_hypernetwork = gr.Button(value="Create", variant='primary')
+
                 with gr.Group():
                     gr.HTML(value="<p style='margin-bottom: 0.7em'>Preprocess images</p>")
 
@@ -1012,6 +999,7 @@ def create_ui(wrap_gradio_gpu_call):
                 with gr.Group():
                     gr.HTML(value="<p style='margin-bottom: 0.7em'>Train an embedding; must specify a directory with a set of 512x512 images</p>")
                     train_embedding_name = gr.Dropdown(label='Embedding', choices=sorted(sd_hijack.model_hijack.embedding_db.word_embeddings.keys()))
+                    train_hypernetwork_name = gr.Dropdown(label='Hypernetwork', choices=[x for x in shared.hypernetworks.keys()])
                     learn_rate = gr.Number(label='Learning rate', value=5.0e-03)
                     dataset_directory = gr.Textbox(label='Dataset directory', placeholder="Path to directory with input images")
                     log_directory = gr.Textbox(label='Log directory', placeholder="Path to directory where to write outputs", value="textual_inversion")
@@ -1019,15 +1007,12 @@ def create_ui(wrap_gradio_gpu_call):
                     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)
+                    preview_image_prompt = gr.Textbox(label='Preview prompt', value="")
 
                     with gr.Row():
-                        with gr.Column(scale=2):
-                            gr.HTML(value="")
-
-                        with gr.Column():
-                            with gr.Row():
-                                interrupt_training = gr.Button(value="Interrupt")
-                                train_embedding = gr.Button(value="Train", variant='primary')
+                        interrupt_training = gr.Button(value="Interrupt")
+                        train_hypernetwork = gr.Button(value="Train Hypernetwork", variant='primary')
+                        train_embedding = gr.Button(value="Train Embedding", variant='primary')
 
             with gr.Column():
                 progressbar = gr.HTML(elem_id="ti_progressbar")
@@ -1053,6 +1038,18 @@ def create_ui(wrap_gradio_gpu_call):
             ]
         )
 
+        create_hypernetwork.click(
+            fn=modules.hypernetwork.ui.create_hypernetwork,
+            inputs=[
+                new_hypernetwork_name,
+            ],
+            outputs=[
+                train_hypernetwork_name,
+                ti_output,
+                ti_outcome,
+            ]
+        )
+
         run_preprocess.click(
             fn=wrap_gradio_gpu_call(modules.textual_inversion.ui.preprocess, extra_outputs=[gr.update()]),
             _js="start_training_textual_inversion",
@@ -1088,12 +1085,33 @@ def create_ui(wrap_gradio_gpu_call):
             ]
         )
 
+        train_hypernetwork.click(
+            fn=wrap_gradio_gpu_call(modules.hypernetwork.ui.train_hypernetwork, extra_outputs=[gr.update()]),
+            _js="start_training_textual_inversion",
+            inputs=[
+                train_hypernetwork_name,
+                learn_rate,
+                dataset_directory,
+                log_directory,
+                steps,
+                create_image_every,
+                save_embedding_every,
+                template_file,
+                preview_image_prompt,
+            ],
+            outputs=[
+                ti_output,
+                ti_outcome,
+            ]
+        )
+
         interrupt_training.click(
             fn=lambda: shared.state.interrupt(),
             inputs=[],
             outputs=[],
         )
 
+
     def create_setting_component(key):
         def fun():
             return opts.data[key] if key in opts.data else opts.data_labels[key].default

scripts/xy_grid.py

@@ -1,8 +1,9 @@
 from collections import namedtuple
 from copy import copy
-from itertools import permutations
+from itertools import permutations, chain
 import random
-
+import csv
+from io import StringIO
 from PIL import Image
 import numpy as np
 
@@ -76,6 +77,10 @@ def apply_checkpoint(p, x, xs):
     modules.sd_models.reload_model_weights(shared.sd_model, info)
 
 
+def apply_hypernetwork(p, x, xs):
+    shared.hypernetwork = shared.hypernetworks.get(x, None)
+
+
 def format_value_add_label(p, opt, x):
     if type(x) == float:
         x = round(x, 8)
@@ -121,6 +126,7 @@ axis_options = [
     AxisOption("Prompt order", str_permutations, apply_order, format_value_join_list),
     AxisOption("Sampler", str, apply_sampler, format_value),
     AxisOption("Checkpoint name", str, apply_checkpoint, format_value),
+    AxisOption("Hypernetwork", str, apply_hypernetwork, format_value),
     AxisOption("Sigma Churn", float, apply_field("s_churn"), format_value_add_label),
     AxisOption("Sigma min", float, apply_field("s_tmin"), format_value_add_label),
     AxisOption("Sigma max", float, apply_field("s_tmax"), format_value_add_label),
@@ -168,7 +174,6 @@ re_range_float = re.compile(r"\s*([+-]?\s*\d+(?:.\d*)?)\s*-\s*([+-]?\s*\d+(?:.\d
 re_range_count = re.compile(r"\s*([+-]?\s*\d+)\s*-\s*([+-]?\s*\d+)(?:\s*\[(\d+)\s*\])?\s*")
 re_range_count_float = re.compile(r"\s*([+-]?\s*\d+(?:.\d*)?)\s*-\s*([+-]?\s*\d+(?:.\d*)?)(?:\s*\[(\d+(?:.\d*)?)\s*\])?\s*")
 
-
 class Script(scripts.Script):
     def title(self):
         return "X/Y plot"
@@ -193,11 +198,13 @@ class Script(scripts.Script):
         modules.processing.fix_seed(p)
         p.batch_size = 1
 
+        initial_hn = shared.hypernetwork
+
         def process_axis(opt, vals):
             if opt.label == 'Nothing':
                 return [0]
 
-            valslist = [x.strip() for x in vals.split(",")]
+            valslist = [x.strip() for x in chain.from_iterable(csv.reader(StringIO(vals)))]
 
             if opt.type == int:
                 valslist_ext = []
@@ -300,4 +307,6 @@ class Script(scripts.Script):
         # restore checkpoint in case it was changed by axes
         modules.sd_models.reload_model_weights(shared.sd_model)
 
+        shared.hypernetwork = initial_hn
+
         return processed

style.css

@@ -404,6 +404,10 @@ input[type="range"]{
 	color: red;
 }
 
+.gallery-item {
+    --tw-bg-opacity: 0 !important;
+}
+
 #img2img_image div.h-60{
     height: 480px;
 }

textual_inversion_templates/hypernetwork.txt

@@ -0,0 +1,27 @@
+a photo of a [filewords]
+a rendering of a [filewords]
+a cropped photo of the [filewords]
+the photo of a [filewords]
+a photo of a clean [filewords]
+a photo of a dirty [filewords]
+a dark photo of the [filewords]
+a photo of my [filewords]
+a photo of the cool [filewords]
+a close-up photo of a [filewords]
+a bright photo of the [filewords]
+a cropped photo of a [filewords]
+a photo of the [filewords]
+a good photo of the [filewords]
+a photo of one [filewords]
+a close-up photo of the [filewords]
+a rendition of the [filewords]
+a photo of the clean [filewords]
+a rendition of a [filewords]
+a photo of a nice [filewords]
+a good photo of a [filewords]
+a photo of the nice [filewords]
+a photo of the small [filewords]
+a photo of the weird [filewords]
+a photo of the large [filewords]
+a photo of a cool [filewords]
+a photo of a small [filewords]

textual_inversion_templates/none.txt

@@ -0,0 +1 @@
+picture

webui.py

@@ -74,6 +74,15 @@ def wrap_gradio_gpu_call(func, extra_outputs=None):
     return modules.ui.wrap_gradio_call(f, extra_outputs=extra_outputs)
 
 
+def set_hypernetwork():
+    shared.hypernetwork = shared.hypernetworks.get(shared.opts.sd_hypernetwork, None)
+
+
+shared.reload_hypernetworks()
+shared.opts.onchange("sd_hypernetwork", set_hypernetwork)
+set_hypernetwork()
+
+
 modules.scripts.load_scripts(os.path.join(script_path, "scripts"))
 
 shared.sd_model = modules.sd_models.load_model()