hypernetwork training mk1

README.md

@@ -16,7 +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 automatically adjust attention to selected text (code contributed by anonymous user)
+    - 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
@@ -65,7 +65,6 @@ Check the [custom scripts](https://github.com/AUTOMATIC1111/stable-diffusion-web
 - [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`
-- No token limit for prompts (original stable diffusion lets you use up to 75 tokens)
 
 ## 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/contextMenus.js

@@ -1,172 +0,0 @@
-
-contextMenuInit = function(){
-  let eventListenerApplied=false;
-  let menuSpecs = new Map();
-
-  const uid = function(){
-    return Date.now().toString(36) + Math.random().toString(36).substr(2);
-  }
-
-  function showContextMenu(event,element,menuEntries){
-    let posx = event.clientX + document.body.scrollLeft + document.documentElement.scrollLeft;
-    let posy = event.clientY + document.body.scrollTop + document.documentElement.scrollTop; 
-
-    let oldMenu = gradioApp().querySelector('#context-menu')
-    if(oldMenu){
-      oldMenu.remove()
-    }
-
-    let tabButton = gradioApp().querySelector('button')
-    let baseStyle = window.getComputedStyle(tabButton)
-
-    const contextMenu = document.createElement('nav')
-    contextMenu.id = "context-menu"
-    contextMenu.style.background = baseStyle.background
-    contextMenu.style.color = baseStyle.color
-    contextMenu.style.fontFamily = baseStyle.fontFamily
-    contextMenu.style.top = posy+'px'
-    contextMenu.style.left = posx+'px'
-
-
-
-    const contextMenuList = document.createElement('ul')
-    contextMenuList.className = 'context-menu-items';
-    contextMenu.append(contextMenuList);
-
-    menuEntries.forEach(function(entry){
-      let contextMenuEntry = document.createElement('a')
-      contextMenuEntry.innerHTML = entry['name']
-      contextMenuEntry.addEventListener("click", function(e) {
-        entry['func']();
-      })
-      contextMenuList.append(contextMenuEntry);
-
-    })
-
-    gradioApp().getRootNode().appendChild(contextMenu)
-
-    let menuWidth = contextMenu.offsetWidth + 4;
-    let menuHeight = contextMenu.offsetHeight + 4;
-
-    let windowWidth = window.innerWidth;
-    let windowHeight = window.innerHeight;
-
-    if ( (windowWidth - posx) < menuWidth ) {
-      contextMenu.style.left = windowWidth - menuWidth + "px";
-    }
-
-    if ( (windowHeight - posy) < menuHeight ) {
-      contextMenu.style.top = windowHeight - menuHeight + "px";
-    }
-
-  }
-
-  function appendContextMenuOption(targetEmementSelector,entryName,entryFunction){
-    
-    currentItems = menuSpecs.get(targetEmementSelector)
-    
-    if(!currentItems){
-      currentItems = []
-      menuSpecs.set(targetEmementSelector,currentItems);
-    }
-    let newItem = {'id':targetEmementSelector+'_'+uid(), 
-                   'name':entryName,
-                   'func':entryFunction,
-                   'isNew':true}
-
-    currentItems.push(newItem)
-    return newItem['id']
-  }
-
-  function removeContextMenuOption(uid){
-    menuSpecs.forEach(function(v,k) {
-      let index = -1
-      v.forEach(function(e,ei){if(e['id']==uid){index=ei}})
-      if(index>=0){
-        v.splice(index, 1);
-      }
-    })
-  }
-
-  function addContextMenuEventListener(){
-    if(eventListenerApplied){
-      return;
-    }
-    gradioApp().addEventListener("click", function(e) {
-      let source = e.composedPath()[0]
-      if(source.id && source.indexOf('check_progress')>-1){
-        return
-      }
-      
-      let oldMenu = gradioApp().querySelector('#context-menu')
-      if(oldMenu){
-        oldMenu.remove()
-      }
-    });
-    gradioApp().addEventListener("contextmenu", function(e) {
-      let oldMenu = gradioApp().querySelector('#context-menu')
-      if(oldMenu){
-        oldMenu.remove()
-      }
-      menuSpecs.forEach(function(v,k) {
-        if(e.composedPath()[0].matches(k)){
-          showContextMenu(e,e.composedPath()[0],v)
-          e.preventDefault()
-          return
-        }
-      })
-    });
-    eventListenerApplied=true
-  
-  }
-
-  return [appendContextMenuOption, removeContextMenuOption, addContextMenuEventListener]
-}
-
-initResponse = contextMenuInit()
-appendContextMenuOption     = initResponse[0]
-removeContextMenuOption     = initResponse[1]
-addContextMenuEventListener = initResponse[2]
-
-
-//Start example Context Menu Items
-generateOnRepeatId = appendContextMenuOption('#txt2img_generate','Generate forever',function(){
-  let genbutton = gradioApp().querySelector('#txt2img_generate');
-  let interruptbutton = gradioApp().querySelector('#txt2img_interrupt');
-  if(!interruptbutton.offsetParent){
-    genbutton.click();
-  }
-  clearInterval(window.generateOnRepeatInterval)
-  window.generateOnRepeatInterval = setInterval(function(){
-    if(!interruptbutton.offsetParent){
-      genbutton.click();
-    }
-  },
-  500)}
-)
-
-cancelGenerateForever = function(){ 
-  clearInterval(window.generateOnRepeatInterval) 
-  let interruptbutton = gradioApp().querySelector('#txt2img_interrupt');
-  if(interruptbutton.offsetParent){
-      interruptbutton.click();
-  }
-}
-
-appendContextMenuOption('#txt2img_interrupt','Cancel generate forever',cancelGenerateForever)
-appendContextMenuOption('#txt2img_generate', 'Cancel generate forever',cancelGenerateForever)
-
-
-appendContextMenuOption('#roll','Roll three',
-  function(){ 
-    let rollbutton = gradioApp().querySelector('#roll');
-    setTimeout(function(){rollbutton.click()},100)
-    setTimeout(function(){rollbutton.click()},200)
-    setTimeout(function(){rollbutton.click()},300)
-  }
-)
-//End example Context Menu Items
-
-onUiUpdate(function(){
-  addContextMenuEventListener()
-});

javascript/edit-attention.js

@@ -1,5 +1,5 @@
 addEventListener('keydown', (event) => {
-	let target = event.originalTarget || event.composedPath()[0];
+	let target = event.originalTarget;
 	if (!target.hasAttribute("placeholder")) return;
 	if (!target.placeholder.toLowerCase().includes("prompt")) return;
 

javascript/hints.js

@@ -35,7 +35,6 @@ titles = {
     "Denoising strength": "Determines how little respect the algorithm should have for image's content. At 0, nothing will change, and at 1 you'll get an unrelated image. With values below 1.0, processing will take less steps than the Sampling Steps slider specifies.",
     "Denoising strength change factor": "In loopback mode, on each loop the denoising strength is multiplied by this value. <1 means decreasing variety so your sequence will converge on a fixed picture. >1 means increasing variety so your sequence will become more and more chaotic.",
 
-    "Skip": "Stop processing current image and continue processing.",
     "Interrupt": "Stop processing images and return any results accumulated so far.",
     "Save": "Write image to a directory (default - log/images) and generation parameters into csv file.",
 

javascript/imageviewer.js

@@ -86,9 +86,6 @@ function showGalleryImage(){
         
         if(fullImg_preview != null){
             fullImg_preview.forEach(function function_name(e) {
-                if (e.dataset.modded)
-                    return;
-                e.dataset.modded = true;
                 if(e && e.parentElement.tagName == 'DIV'){
 
                     e.style.cursor='pointer'

javascript/progressbar.js

@@ -1,9 +1,8 @@
 // code related to showing and updating progressbar shown as the image is being made
 global_progressbars = {}
 
-function check_progressbar(id_part, id_progressbar, id_progressbar_span, id_skip, id_interrupt, id_preview, id_gallery){
+function check_progressbar(id_part, id_progressbar, id_progressbar_span, id_interrupt, id_preview, id_gallery){
     var progressbar = gradioApp().getElementById(id_progressbar)
-    var skip = id_skip ? gradioApp().getElementById(id_skip) : null
     var interrupt = gradioApp().getElementById(id_interrupt)
     
     if(opts.show_progress_in_title && progressbar && progressbar.offsetParent){
@@ -33,37 +32,30 @@ function check_progressbar(id_part, id_progressbar, id_progressbar_span, id_skip
 
                 var progressDiv = gradioApp().querySelectorAll('#' + id_progressbar_span).length > 0;
                 if(!progressDiv){
-                    if (skip) {
-                        skip.style.display = "none"
-                    }
                     interrupt.style.display = "none"
                 }
             }
 
-            window.setTimeout(function() { requestMoreProgress(id_part, id_progressbar_span, id_skip, id_interrupt) }, 500)
+            window.setTimeout(function(){ requestMoreProgress(id_part, id_progressbar_span, id_interrupt) }, 500)
         });
         mutationObserver.observe( progressbar, { childList:true, subtree:true })
 	}
 }
 
 onUiUpdate(function(){
-    check_progressbar('txt2img', 'txt2img_progressbar', 'txt2img_progress_span', 'txt2img_skip', 'txt2img_interrupt', 'txt2img_preview', 'txt2img_gallery')
-    check_progressbar('img2img', 'img2img_progressbar', 'img2img_progress_span', 'img2img_skip', 'img2img_interrupt', 'img2img_preview', 'img2img_gallery')
-    check_progressbar('ti', 'ti_progressbar', 'ti_progress_span', '', 'ti_interrupt', 'ti_preview', 'ti_gallery')
+    check_progressbar('txt2img', 'txt2img_progressbar', 'txt2img_progress_span', 'txt2img_interrupt', 'txt2img_preview', 'txt2img_gallery')
+    check_progressbar('img2img', 'img2img_progressbar', 'img2img_progress_span', 'img2img_interrupt', 'img2img_preview', 'img2img_gallery')
+    check_progressbar('ti', 'ti_progressbar', 'ti_progress_span', 'ti_interrupt', 'ti_preview', 'ti_gallery')
 })
 
-function requestMoreProgress(id_part, id_progressbar_span, id_skip, id_interrupt){
+function requestMoreProgress(id_part, id_progressbar_span, id_interrupt){
     btn = gradioApp().getElementById(id_part+"_check_progress");
     if(btn==null) return;
 
     btn.click();
     var progressDiv = gradioApp().querySelectorAll('#' + id_progressbar_span).length > 0;
-    var skip = id_skip ? gradioApp().getElementById(id_skip) : null
     var interrupt = gradioApp().getElementById(id_interrupt)
     if(progressDiv && interrupt){
-        if (skip) {
-            skip.style.display = "block"
-        }
         interrupt.style.display = "block"
     }
 }

launch.py

@@ -4,7 +4,6 @@ import os
 import sys
 import importlib.util
 import shlex
-import platform
 
 dir_repos = "repositories"
 dir_tmp = "tmp"
@@ -32,7 +31,6 @@ def extract_arg(args, name):
 
 
 args, skip_torch_cuda_test = extract_arg(args, '--skip-torch-cuda-test')
-xformers = '--xformers' in args
 
 
 def repo_dir(name):
@@ -126,12 +124,6 @@ if not is_installed("gfpgan"):
 if not is_installed("clip"):
     run_pip(f"install {clip_package}", "clip")
 
-if not is_installed("xformers") and xformers and platform.python_version().startswith("3.10"):
-    if platform.system() == "Windows":
-        run_pip("install https://github.com/C43H66N12O12S2/stable-diffusion-webui/releases/download/a/xformers-0.0.14.dev0-cp310-cp310-win_amd64.whl", "xformers")
-    elif platform.system() == "Linux":
-        run_pip("install xformers", "xformers")
-
 os.makedirs(dir_repos, exist_ok=True)
 
 git_clone("https://github.com/CompVis/stable-diffusion.git", repo_dir('stable-diffusion'), "Stable Diffusion", stable_diffusion_commit_hash)

modules/esrgan_model.py

@@ -111,7 +111,7 @@ class UpscalerESRGAN(Upscaler):
             print("Unable to load %s from %s" % (self.model_path, filename))
             return None
 
-        pretrained_net = torch.load(filename, map_location='cpu' if devices.device_esrgan.type == 'mps' else None)
+        pretrained_net = torch.load(filename, map_location='cpu' if shared.device.type == 'mps' else None)
         crt_model = arch.RRDBNet(3, 3, 64, 23, gc=32)
 
         pretrained_net = fix_model_layers(crt_model, pretrained_net)

modules/hypernetwork.py

@@ -1,88 +0,0 @@
-import glob
-import os
-import sys
-import traceback
-
-import torch
-
-from ldm.util import default
-from modules import devices, shared
-import torch
-from torch import einsum
-from einops import rearrange, repeat
-
-
-class HypernetworkModule(torch.nn.Module):
-    def __init__(self, dim, state_dict):
-        super().__init__()
-
-        self.linear1 = torch.nn.Linear(dim, dim * 2)
-        self.linear2 = torch.nn.Linear(dim * 2, dim)
-
-        self.load_state_dict(state_dict, strict=True)
-        self.to(devices.device)
-
-    def forward(self, x):
-        return x + (self.linear2(self.linear1(x)))
-
-
-class Hypernetwork:
-    filename = None
-    name = None
-
-    def __init__(self, filename):
-        self.filename = filename
-        self.name = os.path.splitext(os.path.basename(filename))[0]
-        self.layers = {}
-
-        state_dict = torch.load(filename, map_location='cpu')
-        for size, sd in state_dict.items():
-            self.layers[size] = (HypernetworkModule(size, sd[0]), HypernetworkModule(size, sd[1]))
-
-
-def load_hypernetworks(path):
-    res = {}
-
-    for filename in glob.iglob(os.path.join(path, '**/*.pt'), recursive=True):
-        try:
-            hn = Hypernetwork(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 = shared.selected_hypernetwork()
-    hypernetwork_layers = (hypernetwork.layers if hypernetwork is not None else {}).get(context.shape[2], None)
-
-    if hypernetwork_layers is not None:
-        k = self.to_k(hypernetwork_layers[0](context))
-        v = self.to_v(hypernetwork_layers[1](context))
-    else:
-        k = self.to_k(context)
-        v = self.to_v(context)
-
-    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)

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/img2img.py

@@ -32,8 +32,6 @@ def process_batch(p, input_dir, output_dir, args):
 
     for i, image in enumerate(images):
         state.job = f"{i+1} out of {len(images)}"
-        if state.skipped:
-            state.skipped = False
 
         if state.interrupted:
             break

modules/processing.py

@@ -141,7 +141,6 @@ class Processed:
         self.all_subseeds = all_subseeds or [self.subseed]
         self.infotexts = infotexts or [info]
 
-
     def js(self):
         obj = {
             "prompt": self.prompt,
@@ -313,7 +312,6 @@ def process_images(p: StableDiffusionProcessing) -> Processed:
         os.makedirs(p.outpath_grids, exist_ok=True)
 
     modules.sd_hijack.model_hijack.apply_circular(p.tiling)
-    modules.sd_hijack.model_hijack.clear_comments()
 
     comments = {}
 
@@ -351,9 +349,6 @@ def process_images(p: StableDiffusionProcessing) -> Processed:
             state.job_count = p.n_iter
 
         for n in range(p.n_iter):
-            if state.skipped:
-                state.skipped = False
-            
             if state.interrupted:
                 break
 
@@ -380,7 +375,7 @@ def process_images(p: StableDiffusionProcessing) -> Processed:
             with devices.autocast():
                 samples_ddim = p.sample(conditioning=c, unconditional_conditioning=uc, seeds=seeds, subseeds=subseeds, subseed_strength=p.subseed_strength)
 
-            if state.interrupted or state.skipped:
+            if state.interrupted:
 
                 # if we are interruped, sample returns just noise
                 # use the image collected previously in sampler loop

modules/prompt_parser.py

@@ -13,14 +13,13 @@ import lark
 
 schedule_parser = lark.Lark(r"""
 !start: (prompt | /[][():]/+)*
-prompt: (emphasized | scheduled | alternate | plain | WHITESPACE)*
+prompt: (emphasized | scheduled | plain | WHITESPACE)*
 !emphasized: "(" prompt ")"
         | "(" prompt ":" prompt ")"
         | "[" prompt "]"
 scheduled: "[" [prompt ":"] prompt ":" [WHITESPACE] NUMBER "]"
-alternate: "[" prompt ("|" prompt)+ "]"
 WHITESPACE: /\s+/
-plain: /([^\\\[\]():|]|\\.)+/
+plain: /([^\\\[\]():]|\\.)+/
 %import common.SIGNED_NUMBER -> NUMBER
 """)
 
@@ -60,8 +59,6 @@ def get_learned_conditioning_prompt_schedules(prompts, steps):
                     tree.children[-1] *= steps
                 tree.children[-1] = min(steps, int(tree.children[-1]))
                 l.append(tree.children[-1])
-            def alternate(self, tree):
-                l.extend(range(1, steps+1))
         CollectSteps().visit(tree)
         return sorted(set(l))
 
@@ -70,8 +67,6 @@ def get_learned_conditioning_prompt_schedules(prompts, steps):
             def scheduled(self, args):
                 before, after, _, when = args
                 yield before or () if step <= when else after
-            def alternate(self, args):
-                yield next(args[(step - 1)%len(args)])
             def start(self, args):
                 def flatten(x):
                     if type(x) == str:
@@ -244,15 +239,6 @@ def reconstruct_multicond_batch(c: MulticondLearnedConditioning, current_step):
 
         conds_list.append(conds_for_batch)
 
-    # if prompts have wildly different lengths above the limit we'll get tensors fo different shapes
-    # and won't be able to torch.stack them. So this fixes that.
-    token_count = max([x.shape[0] for x in tensors])
-    for i in range(len(tensors)):
-        if tensors[i].shape[0] != token_count:
-            last_vector = tensors[i][-1:]
-            last_vector_repeated = last_vector.repeat([token_count - tensors[i].shape[0], 1])
-            tensors[i] = torch.vstack([tensors[i], last_vector_repeated])
-
     return conds_list, torch.stack(tensors).to(device=param.device, dtype=param.dtype)
 
 

modules/sd_hijack.py

@@ -8,7 +8,7 @@ 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, hypernetwork
+from modules import prompt_parser, devices, sd_hijack_optimizations, shared
 from modules.shared import opts, device, cmd_opts
 
 import ldm.modules.attention
@@ -18,37 +18,27 @@ 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
 
+
 def apply_optimizations():
     undo_optimizations()
 
     ldm.modules.diffusionmodules.model.nonlinearity = silu
 
-    if cmd_opts.force_enable_xformers or (cmd_opts.xformers and shared.xformers_available and torch.version.cuda and torch.cuda.get_device_capability(shared.device) == (8, 6)):
-        print("Applying xformers cross attention optimization.")
-        ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.xformers_attention_forward
-        ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.xformers_attnblock_forward
-    elif cmd_opts.opt_split_attention_v1:
-        print("Applying v1 cross attention optimization.")
+    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()):
-        print("Applying cross attention optimization.")
         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward
         ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.cross_attention_attnblock_forward
 
 
 def undo_optimizations():
+    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
 
 
-def get_target_prompt_token_count(token_count):
-    if token_count < 75:
-        return 75
-
-    return math.ceil(token_count / 10) * 10
-
-
 class StableDiffusionModelHijack:
     fixes = None
     comments = []
@@ -94,12 +84,10 @@ class StableDiffusionModelHijack:
         for layer in [layer for layer in self.layers if type(layer) == torch.nn.Conv2d]:
             layer.padding_mode = 'circular' if enable else 'zeros'
 
-    def clear_comments(self):
-        self.comments = []
-
     def tokenize(self, text):
+        max_length = self.clip.max_length - 2
         _, remade_batch_tokens, _, _, _, token_count = self.clip.process_text([text])
-        return remade_batch_tokens[0], token_count, get_target_prompt_token_count(token_count)
+        return remade_batch_tokens[0], token_count, max_length
 
 
 class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
@@ -108,6 +96,7 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
         self.wrapped = wrapped
         self.hijack: StableDiffusionModelHijack = hijack
         self.tokenizer = wrapped.tokenizer
+        self.max_length = wrapped.max_length
         self.token_mults = {}
 
         tokens_with_parens = [(k, v) for k, v in self.tokenizer.get_vocab().items() if '(' in k or ')' in k or '[' in k or ']' in k]
@@ -129,6 +118,7 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
     def tokenize_line(self, line, used_custom_terms, hijack_comments):
         id_start = self.wrapped.tokenizer.bos_token_id
         id_end = self.wrapped.tokenizer.eos_token_id
+        maxlen = self.wrapped.max_length
 
         if opts.enable_emphasis:
             parsed = prompt_parser.parse_prompt_attention(line)
@@ -160,12 +150,19 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
                     used_custom_terms.append((embedding.name, embedding.checksum()))
                     i += embedding_length_in_tokens
 
-        token_count = len(remade_tokens)
-        prompt_target_length = get_target_prompt_token_count(token_count)
-        tokens_to_add = prompt_target_length - len(remade_tokens) + 1
+        if len(remade_tokens) > maxlen - 2:
+            vocab = {v: k for k, v in self.wrapped.tokenizer.get_vocab().items()}
+            ovf = remade_tokens[maxlen - 2:]
+            overflowing_words = [vocab.get(int(x), "") for x in ovf]
+            overflowing_text = self.wrapped.tokenizer.convert_tokens_to_string(''.join(overflowing_words))
+            hijack_comments.append(f"Warning: too many input tokens; some ({len(overflowing_words)}) have been truncated:\n{overflowing_text}\n")
 
-        remade_tokens = [id_start] + remade_tokens + [id_end] * tokens_to_add
-        multipliers = [1.0] + multipliers + [1.0] * tokens_to_add
+        token_count = len(remade_tokens)
+        remade_tokens = remade_tokens + [id_end] * (maxlen - 2 - len(remade_tokens))
+        remade_tokens = [id_start] + remade_tokens[0:maxlen - 2] + [id_end]
+
+        multipliers = multipliers + [1.0] * (maxlen - 2 - len(multipliers))
+        multipliers = [1.0] + multipliers[0:maxlen - 2] + [1.0]
 
         return remade_tokens, fixes, multipliers, token_count
 
@@ -182,8 +179,7 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
             if line in cache:
                 remade_tokens, fixes, multipliers = cache[line]
             else:
-                remade_tokens, fixes, multipliers, current_token_count = self.tokenize_line(line, used_custom_terms, hijack_comments)
-                token_count = max(current_token_count, token_count)
+                remade_tokens, fixes, multipliers, token_count = self.tokenize_line(line, used_custom_terms, hijack_comments)
 
                 cache[line] = (remade_tokens, fixes, multipliers)
 
@@ -197,7 +193,7 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
     def process_text_old(self, text):
         id_start = self.wrapped.tokenizer.bos_token_id
         id_end = self.wrapped.tokenizer.eos_token_id
-        maxlen = self.wrapped.max_length  # you get to stay at 77
+        maxlen = self.wrapped.max_length
         used_custom_terms = []
         remade_batch_tokens = []
         overflowing_words = []
@@ -269,24 +265,17 @@ class FrozenCLIPEmbedderWithCustomWords(torch.nn.Module):
             batch_multipliers, remade_batch_tokens, used_custom_terms, hijack_comments, hijack_fixes, token_count = self.process_text(text)
 
         self.hijack.fixes = hijack_fixes
-        self.hijack.comments += hijack_comments
+        self.hijack.comments = hijack_comments
 
         if len(used_custom_terms) > 0:
             self.hijack.comments.append("Used embeddings: " + ", ".join([f'{word} [{checksum}]' for word, checksum in used_custom_terms]))
 
-        target_token_count = get_target_prompt_token_count(token_count) + 2
-
-        position_ids_array = [min(x, 75) for x in range(target_token_count-1)] + [76]
-        position_ids = torch.asarray(position_ids_array, device=devices.device).expand((1, -1))
-
-        remade_batch_tokens_of_same_length = [x + [self.wrapped.tokenizer.eos_token_id] * (target_token_count - len(x)) for x in remade_batch_tokens]
-        tokens = torch.asarray(remade_batch_tokens_of_same_length).to(device)
-        outputs = self.wrapped.transformer(input_ids=tokens, position_ids=position_ids)
+        tokens = torch.asarray(remade_batch_tokens).to(device)
+        outputs = self.wrapped.transformer(input_ids=tokens)
         z = outputs.last_hidden_state
 
         # restoring original mean is likely not correct, but it seems to work well to prevent artifacts that happen otherwise
-        batch_multipliers_of_same_length = [x + [1.0] * (target_token_count - len(x)) for x in batch_multipliers]
-        batch_multipliers = torch.asarray(batch_multipliers_of_same_length).to(device)
+        batch_multipliers = torch.asarray(batch_multipliers).to(device)
         original_mean = z.mean()
         z *= batch_multipliers.reshape(batch_multipliers.shape + (1,)).expand(z.shape)
         new_mean = z.mean()

modules/sd_hijack_optimizations.py

@@ -1,7 +1,4 @@
 import math
-import sys
-import traceback
-
 import torch
 from torch import einsum
 
@@ -10,37 +7,18 @@ from einops import rearrange
 
 from modules import shared
 
-if shared.cmd_opts.xformers:
-    try:
-        import xformers.ops
-        import functorch
-        xformers._is_functorch_available = True
-        shared.xformers_available = True
-    except Exception:
-        print("Cannot import xformers", file=sys.stderr)
-        print(traceback.format_exc(), file=sys.stderr)
-
 
 # see https://github.com/basujindal/stable-diffusion/pull/117 for discussion
 def split_cross_attention_forward_v1(self, x, context=None, mask=None):
     h = self.heads
 
-    q_in = self.to_q(x)
+    q = self.to_q(x)
     context = default(context, x)
-
-    hypernetwork = shared.selected_hypernetwork()
-    hypernetwork_layers = (hypernetwork.layers if 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 = self.to_k(context)
+    v = self.to_v(context)
     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))
-    del q_in, k_in, v_in
+    q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
 
     r1 = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device)
     for i in range(0, q.shape[0], 2):
@@ -53,7 +31,6 @@ def split_cross_attention_forward_v1(self, x, context=None, mask=None):
 
         r1[i:end] = einsum('b i j, b j d -> b i d', s2, v[i:end])
         del s2
-    del q, k, v
 
     r2 = rearrange(r1, '(b h) n d -> b n (h d)', h=h)
     del r1
@@ -68,8 +45,7 @@ def split_cross_attention_forward(self, x, context=None, mask=None):
     q_in = self.to_q(x)
     context = default(context, x)
 
-    hypernetwork = shared.selected_hypernetwork()
-    hypernetwork_layers = (hypernetwork.layers if hypernetwork is not None else {}).get(context.shape[2], None)
+    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))
@@ -128,25 +104,6 @@ def split_cross_attention_forward(self, x, context=None, mask=None):
 
     return self.to_out(r2)
 
-def xformers_attention_forward(self, x, context=None, mask=None):
-    h = self.heads
-    q_in = self.to_q(x)
-    context = default(context, x)
-    hypernetwork = shared.selected_hypernetwork()
-    hypernetwork_layers = (hypernetwork.layers if 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)
-    q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b n h d', h=h), (q_in, k_in, v_in))
-    del q_in, k_in, v_in
-    out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None)
-
-    out = rearrange(out, 'b n h d -> b n (h d)', h=h)
-    return self.to_out(out)
-
 def cross_attention_attnblock_forward(self, x):
         h_ = x
         h_ = self.norm(h_)
@@ -209,16 +166,3 @@ def cross_attention_attnblock_forward(self, x):
         h3 += x
 
         return h3
-    
-def xformers_attnblock_forward(self, x):
-    try:
-        h_ = x
-        h_ = self.norm(h_)
-        q1 = self.q(h_).contiguous()
-        k1 = self.k(h_).contiguous()
-        v = self.v(h_).contiguous()
-        out = xformers.ops.memory_efficient_attention(q1, k1, v)
-        out = self.proj_out(out)
-        return x + out
-    except NotImplementedError:
-        return cross_attention_attnblock_forward(self, x)

modules/sd_models.py

@@ -122,11 +122,7 @@ def load_model_weights(model, checkpoint_file, sd_model_hash):
     pl_sd = torch.load(checkpoint_file, map_location="cpu")
     if "global_step" in pl_sd:
         print(f"Global Step: {pl_sd['global_step']}")
-    
-    if "state_dict" in pl_sd:
-        sd = pl_sd["state_dict"]
-    else:
-        sd = pl_sd
+    sd = pl_sd["state_dict"]
 
     model.load_state_dict(sd, strict=False)
 

modules/sd_samplers.py

@@ -106,7 +106,7 @@ def extended_tdqm(sequence, *args, desc=None, **kwargs):
     seq = sequence if cmd_opts.disable_console_progressbars else tqdm.tqdm(sequence, *args, desc=state.job, file=shared.progress_print_out, **kwargs)
 
     for x in seq:
-        if state.interrupted or state.skipped:
+        if state.interrupted:
             break
 
         yield x
@@ -142,16 +142,6 @@ class VanillaStableDiffusionSampler:
         assert all([len(conds) == 1 for conds in conds_list]), 'composition via AND is not supported for DDIM/PLMS samplers'
         cond = tensor
 
-        # for DDIM, shapes must match, we can't just process cond and uncond independently;
-        # filling unconditional_conditioning with repeats of the last vector to match length is
-        # not 100% correct but should work well enough
-        if unconditional_conditioning.shape[1] < cond.shape[1]:
-            last_vector = unconditional_conditioning[:, -1:]
-            last_vector_repeated = last_vector.repeat([1, cond.shape[1] - unconditional_conditioning.shape[1], 1])
-            unconditional_conditioning = torch.hstack([unconditional_conditioning, last_vector_repeated])
-        elif unconditional_conditioning.shape[1] > cond.shape[1]:
-            unconditional_conditioning = unconditional_conditioning[:, :cond.shape[1]]
-
         if self.mask is not None:
             img_orig = self.sampler.model.q_sample(self.init_latent, ts)
             x_dec = img_orig * self.mask + self.nmask * x_dec
@@ -231,29 +221,18 @@ class CFGDenoiser(torch.nn.Module):
 
         x_in = torch.cat([torch.stack([x[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [x])
         sigma_in = torch.cat([torch.stack([sigma[i] for _ in range(n)]) for i, n in enumerate(repeats)] + [sigma])
+        cond_in = torch.cat([tensor, uncond])
 
-        if tensor.shape[1] == uncond.shape[1]:
-            cond_in = torch.cat([tensor, uncond])
-
-            if shared.batch_cond_uncond:
-                x_out = self.inner_model(x_in, sigma_in, cond=cond_in)
-            else:
-                x_out = torch.zeros_like(x_in)
-                for batch_offset in range(0, x_out.shape[0], batch_size):
-                    a = batch_offset
-                    b = a + batch_size
-                    x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond=cond_in[a:b])
+        if shared.batch_cond_uncond:
+            x_out = self.inner_model(x_in, sigma_in, cond=cond_in)
         else:
             x_out = torch.zeros_like(x_in)
-            batch_size = batch_size*2 if shared.batch_cond_uncond else batch_size
-            for batch_offset in range(0, tensor.shape[0], batch_size):
+            for batch_offset in range(0, x_out.shape[0], batch_size):
                 a = batch_offset
-                b = min(a + batch_size, tensor.shape[0])
-                x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond=tensor[a:b])
+                b = a + batch_size
+                x_out[a:b] = self.inner_model(x_in[a:b], sigma_in[a:b], cond=cond_in[a:b])
 
-            x_out[-uncond.shape[0]:] = self.inner_model(x_in[-uncond.shape[0]:], sigma_in[-uncond.shape[0]:], cond=uncond)
-
-        denoised_uncond = x_out[-uncond.shape[0]:]
+        denoised_uncond = x_out[-batch_size:]
         denoised = torch.clone(denoised_uncond)
 
         for i, conds in enumerate(conds_list):
@@ -275,7 +254,7 @@ def extended_trange(sampler, count, *args, **kwargs):
     seq = range(count) if cmd_opts.disable_console_progressbars else tqdm.trange(count, *args, desc=state.job, file=shared.progress_print_out, **kwargs)
 
     for x in seq:
-        if state.interrupted or state.skipped:
+        if state.interrupted:
             break
 
         if sampler.stop_at is not None and x > sampler.stop_at:

modules/shared.py

@@ -13,7 +13,7 @@ import modules.memmon
 import modules.sd_models
 import modules.styles
 import modules.devices as devices
-from modules import sd_samplers, hypernetwork
+from modules import sd_samplers
 from modules.paths import models_path, script_path, sd_path
 
 sd_model_file = os.path.join(script_path, 'model.ckpt')
@@ -28,6 +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")
@@ -43,8 +44,6 @@ parser.add_argument("--realesrgan-models-path", type=str, help="Path to director
 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("--xformers", action='store_true', help="enable xformers for cross attention layers")
-parser.add_argument("--force-enable-xformers", action='store_true', help="enable xformers for cross attention layers regardless of whether the checking code thinks you can run it; do not make bug reports if this fails to work")
 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")
@@ -75,18 +74,21 @@ device = devices.device
 
 batch_cond_uncond = cmd_opts.always_batch_cond_uncond or not (cmd_opts.lowvram or cmd_opts.medvram)
 parallel_processing_allowed = not cmd_opts.lowvram and not cmd_opts.medvram
-xformers_available = False
+
 config_filename = cmd_opts.ui_settings_file
 
-hypernetworks = hypernetwork.load_hypernetworks(os.path.join(models_path, 'hypernetworks'))
+
+def reload_hypernetworks():
+    from modules.hypernetwork import hypernetwork
+    hypernetworks.clear()
+    hypernetworks.update(hypernetwork.load_hypernetworks(cmd_opts.hypernetwork_dir))
 
 
-def selected_hypernetwork():
-    return hypernetworks.get(opts.sd_hypernetwork, None)
+hypernetworks = {}
+hypernetwork = None
 
 
 class State:
-    skipped = False
     interrupted = False
     job = ""
     job_no = 0
@@ -99,9 +101,6 @@ class State:
     current_image_sampling_step = 0
     textinfo = None
 
-    def skip(self):
-        self.skipped = True
-
     def interrupt(self):
         self.interrupted = True
 
@@ -124,6 +123,8 @@ prompt_styles = modules.styles.StyleDatabase(styles_filename)
 interrogator = modules.interrogate.InterrogateModels("interrogate")
 
 face_restorers = []
+# This was moved to webui.py with the other model "setup" calls.
+# modules.sd_models.list_models()
 
 
 def realesrgan_models_names():

modules/textual_inversion/ui.py

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

modules/ui.py

@@ -37,6 +37,7 @@ import modules.generation_parameters_copypaste
 from modules import prompt_parser
 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()
@@ -191,7 +192,6 @@ def wrap_gradio_call(func, extra_outputs=None):
         # last item is always HTML
         res[-1] += f"<div class='performance'><p class='time'>Time taken: <wbr>{elapsed_text}</p>{vram_html}</div>"
 
-        shared.state.skipped = False
         shared.state.interrupted = False
         shared.state.job_count = 0
 
@@ -412,16 +412,9 @@ def create_toprow(is_img2img):
 
         with gr.Column(scale=1):
             with gr.Row():
-                skip = gr.Button('Skip', elem_id=f"{id_part}_skip")
                 interrupt = gr.Button('Interrupt', elem_id=f"{id_part}_interrupt")
                 submit = gr.Button('Generate', elem_id=f"{id_part}_generate", variant='primary')
 
-                skip.click(
-                    fn=lambda: shared.state.skip(),
-                    inputs=[],
-                    outputs=[],
-                )
-
                 interrupt.click(
                     fn=lambda: shared.state.interrupt(),
                     inputs=[],
@@ -946,7 +939,7 @@ def create_ui(wrap_gradio_gpu_call):
                 custom_name = gr.Textbox(label="Custom Name (Optional)")
                 interp_amount = gr.Slider(minimum=0.0, maximum=1.0, step=0.05, label='Interpolation Amount', value=0.3)
                 interp_method = gr.Radio(choices=["Weighted Sum", "Sigmoid", "Inverse Sigmoid"], value="Weighted Sum", label="Interpolation Method")
-                save_as_half = gr.Checkbox(value=False, label="Save as float16")
+                save_as_half = gr.Checkbox(value=False, label="Safe as float16")
                 modelmerger_merge = gr.Button(elem_id="modelmerger_merge", label="Merge", variant='primary')
 
             with gr.Column(variant='panel'):
@@ -973,6 +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>")
 
@@ -980,9 +985,9 @@ def create_ui(wrap_gradio_gpu_call):
                     process_dst = gr.Textbox(label='Destination directory')
 
                     with gr.Row():
-                        process_flip = gr.Checkbox(label='Create flipped copies')
-                        process_split = gr.Checkbox(label='Split oversized images into two')
-                        process_caption = gr.Checkbox(label='Use BLIP caption as filename')
+                        process_flip = gr.Checkbox(label='Flip')
+                        process_split = gr.Checkbox(label='Split into two')
+                        process_caption = gr.Checkbox(label='Add caption')
 
                     with gr.Row():
                         with gr.Column(scale=3):
@@ -994,6 +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")
@@ -1001,15 +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")
@@ -1035,6 +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",
@@ -1070,12 +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

requirements.txt

@@ -23,4 +23,3 @@ resize-right
 torchdiffeq
 kornia
 lark
-functorch

requirements_versions.txt

@@ -22,4 +22,3 @@ resize-right==0.0.2
 torchdiffeq==0.2.3
 kornia==0.6.7
 lark==1.1.2
-functorch==0.2.1

scripts/xy_grid.py

@@ -78,8 +78,7 @@ def apply_checkpoint(p, x, xs):
 
 
 def apply_hypernetwork(p, x, xs):
-    hn = shared.hypernetworks.get(x, None)
-    opts.data["sd_hypernetwork"] = hn.name if hn is not None else 'None'
+    shared.hypernetwork = shared.hypernetworks.get(x, None)
 
 
 def format_value_add_label(p, opt, x):
@@ -199,7 +198,7 @@ class Script(scripts.Script):
         modules.processing.fix_seed(p)
         p.batch_size = 1
 
-        initial_hn = opts.sd_hypernetwork
+        initial_hn = shared.hypernetwork
 
         def process_axis(opt, vals):
             if opt.label == 'Nothing':
@@ -308,6 +307,6 @@ class Script(scripts.Script):
         # restore checkpoint in case it was changed by axes
         modules.sd_models.reload_model_weights(shared.sd_model)
 
-        opts.data["sd_hypernetwork"] = initial_hn
+        shared.hypernetwork = initial_hn
 
         return processed

style.css

@@ -393,20 +393,10 @@ input[type="range"]{
 
 #txt2img_interrupt, #img2img_interrupt{
   position: absolute;
-  width: 50%;
+  width: 100%;
   height: 72px;
   background: #b4c0cc;
-  border-radius: 0px;
-  display: none;
-}
-
-#txt2img_skip, #img2img_skip{
-  position: absolute;
-  width: 50%;
-  right: 0px;
-  height: 72px;
-  background: #b4c0cc;
-  border-radius: 0px;
+  border-radius: 8px;
   display: none;
 }
 
@@ -420,31 +410,4 @@ input[type="range"]{
 
 #img2img_image div.h-60{
     height: 480px;
-}
-
-#context-menu{
-    z-index:9999;
-    position:absolute;
-    display:block;
-    padding:0px 0;
-    border:2px solid #a55000;
-    border-radius:8px;
-    box-shadow:1px 1px 2px #CE6400;
-    width: 200px;
-}
-
-.context-menu-items{
-    list-style: none;
-    margin: 0;
-    padding: 0;
-}
-
-.context-menu-items a{
-    display:block;
-    padding:5px;
-    cursor:pointer;
-}
-
-.context-menu-items a:hover{
-    background: #a55000;
-}
+}

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

@@ -5,8 +5,6 @@ import importlib
 import signal
 import threading
 
-from fastapi.middleware.gzip import GZipMiddleware
-
 from modules.paths import script_path
 
 from modules import devices, sd_samplers
@@ -60,7 +58,6 @@ def wrap_gradio_gpu_call(func, extra_outputs=None):
         shared.state.current_latent = None
         shared.state.current_image = None
         shared.state.current_image_sampling_step = 0
-        shared.state.skipped = False
         shared.state.interrupted = False
         shared.state.textinfo = None
 
@@ -77,6 +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()
@@ -95,7 +101,7 @@ def webui():
 
         demo = modules.ui.create_ui(wrap_gradio_gpu_call=wrap_gradio_gpu_call)
         
-        app,local_url,share_url = demo.launch(
+        demo.launch(
             share=cmd_opts.share,
             server_name="0.0.0.0" if cmd_opts.listen else None,
             server_port=cmd_opts.port,
@@ -104,8 +110,6 @@ def webui():
             inbrowser=cmd_opts.autolaunch,
             prevent_thread_lock=True
         )
-        
-        app.add_middleware(GZipMiddleware,minimum_size=1000)
 
         while 1:
             time.sleep(0.5)