Merge branch 'master' of github.com:comfyanonymous/ComfyUI

.ci/update_windows/update.py

@@ -1,6 +1,9 @@
 import pygit2
 from datetime import datetime
 import sys
+import os
+import shutil
+import filecmp
 
 def pull(repo, remote_name='origin', branch='master'):
     for remote in repo.remotes:
@@ -42,7 +45,8 @@ def pull(repo, remote_name='origin', branch='master'):
                 raise AssertionError('Unknown merge analysis result')
 
 pygit2.option(pygit2.GIT_OPT_SET_OWNER_VALIDATION, 0)
-repo = pygit2.Repository(str(sys.argv[1]))
+repo_path = str(sys.argv[1])
+repo = pygit2.Repository(repo_path)
 ident = pygit2.Signature('comfyui', 'comfy@ui')
 try:
     print("stashing current changes")
@@ -51,7 +55,10 @@ except KeyError:
     print("nothing to stash")
 backup_branch_name = 'backup_branch_{}'.format(datetime.today().strftime('%Y-%m-%d_%H_%M_%S'))
 print("creating backup branch: {}".format(backup_branch_name))
-repo.branches.local.create(backup_branch_name, repo.head.peel())
+try:
+    repo.branches.local.create(backup_branch_name, repo.head.peel())
+except:
+    pass
 
 print("checking out master branch")
 branch = repo.lookup_branch('master')
@@ -63,3 +70,41 @@ pull(repo)
 
 print("Done!")
 
+self_update = True
+if len(sys.argv) > 2:
+    self_update = '--skip_self_update' not in sys.argv
+
+update_py_path = os.path.realpath(__file__)
+repo_update_py_path = os.path.join(repo_path, ".ci/update_windows/update.py")
+
+cur_path = os.path.dirname(update_py_path)
+
+
+req_path = os.path.join(cur_path, "current_requirements.txt")
+repo_req_path = os.path.join(repo_path, "requirements.txt")
+
+
+def files_equal(file1, file2):
+    try:
+        return filecmp.cmp(file1, file2, shallow=False)
+    except:
+        return False
+
+def file_size(f):
+    try:
+        return os.path.getsize(f)
+    except:
+        return 0
+
+
+if self_update and not files_equal(update_py_path, repo_update_py_path) and file_size(repo_update_py_path) > 10:
+    shutil.copy(repo_update_py_path, os.path.join(cur_path, "update_new.py"))
+    exit()
+
+if not os.path.exists(req_path) or not files_equal(repo_req_path, req_path):
+    import subprocess
+    try:
+        subprocess.check_call([sys.executable, '-m', 'pip', 'install', '-r', repo_req_path])
+        shutil.copy(repo_req_path, req_path)
+    except:
+        pass

.ci/update_windows/update_comfyui.bat

@@ -1,2 +1,8 @@
+@echo off
 ..\python_embeded\python.exe .\update.py ..\ComfyUI\
-pause
+if exist update_new.py (
+  move /y update_new.py update.py
+  echo Running updater again since it got updated.
+  ..\python_embeded\python.exe .\update.py ..\ComfyUI\ --skip_self_update
+)
+if "%~1"=="" pause

.ci/update_windows/update_comfyui_and_python_dependencies.bat

@@ -1,3 +0,0 @@
-..\python_embeded\python.exe .\update.py ..\ComfyUI\
-..\python_embeded\python.exe -s -m pip install --upgrade torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117 xformers -r ../ComfyUI/requirements.txt pygit2
-pause

.ci/update_windows_cu118/update_comfyui_and_python_dependencies.bat

@@ -1,11 +0,0 @@
-@echo off
-..\python_embeded\python.exe .\update.py ..\ComfyUI\
-echo
-echo This will try to update pytorch and all python dependencies, if you get an error wait for pytorch/xformers to fix their stuff
-echo You should not be running this anyways unless you really have to
-echo
-echo If you just want to update normally, close this and run update_comfyui.bat instead.
-echo
-pause
-..\python_embeded\python.exe -s -m pip install --upgrade torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu118 xformers -r ../ComfyUI/requirements.txt pygit2
-pause

.github/workflows/windows_release_cu118_dependencies.yml

@@ -1,71 +0,0 @@
-name: "Windows Release cu118 dependencies"
-
-on:
-  workflow_dispatch:
-#  push:
-#    branches:
-#      - master
-
-jobs:
-  build_dependencies:
-    env:
-        # you need at least cuda 5.0 for some of the stuff compiled here.
-        TORCH_CUDA_ARCH_LIST: "5.0+PTX 6.0 6.1 7.0 7.5 8.0 8.6 8.9"
-        FORCE_CUDA: 1
-        MAX_JOBS: 1 # will crash otherwise
-        DISTUTILS_USE_SDK: 1 # otherwise distutils will complain on windows about multiple versions of msvc
-        XFORMERS_BUILD_TYPE: "Release"
-    runs-on: windows-latest
-    steps:
-        - name: Cache Built Dependencies
-          uses: actions/cache@v3
-          id: cache-cu118_python_stuff
-          with:
-            path: cu118_python_deps.tar
-            key: ${{ runner.os }}-build-cu118
-
-        - if: steps.cache-cu118_python_stuff.outputs.cache-hit != 'true'
-          uses: actions/checkout@v3
-
-        - if: steps.cache-cu118_python_stuff.outputs.cache-hit != 'true'
-          uses: actions/setup-python@v4
-          with:
-            python-version: '3.10.9'
-
-        - if: steps.cache-cu118_python_stuff.outputs.cache-hit != 'true'
-          uses: comfyanonymous/cuda-toolkit@test
-          id: cuda-toolkit
-          with:
-            cuda: '11.8.0'
-        # copied from xformers github
-        - name: Setup MSVC
-          uses: ilammy/msvc-dev-cmd@v1
-        - name: Configure Pagefile
-          # windows runners will OOM with many CUDA architectures
-          # we cheat here with a page file
-          uses: al-cheb/configure-pagefile-action@v1.3
-          with:
-            minimum-size: 2GB
-        # really unfortunate: https://github.com/ilammy/msvc-dev-cmd#name-conflicts-with-shell-bash
-        - name: Remove link.exe
-          shell: bash
-          run: rm /usr/bin/link
-
-        - if: steps.cache-cu118_python_stuff.outputs.cache-hit != 'true'
-          shell: bash
-          run: |
-            python -m pip wheel --no-cache-dir torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu118 -r requirements.txt pygit2 -w ./temp_wheel_dir
-            python -m pip install --no-cache-dir ./temp_wheel_dir/*
-            echo installed basic
-            git clone --recurse-submodules https://github.com/facebookresearch/xformers.git
-            cd xformers
-            python -m pip install --no-cache-dir wheel setuptools twine
-            echo building xformers
-            python setup.py bdist_wheel -d ../temp_wheel_dir/
-            cd ..
-            rm -rf xformers
-            ls -lah temp_wheel_dir
-            mv temp_wheel_dir cu118_python_deps
-            tar cf cu118_python_deps.tar cu118_python_deps
-
-

.github/workflows/windows_release_cu118_dependencies_2.yml

@@ -1,37 +0,0 @@
-name: "Windows Release cu118 dependencies 2"
-
-on:
-  workflow_dispatch:
-    inputs:
-      xformers:
-        description: 'xformers version'
-        required: true
-        type: string
-        default: "xformers"
-
-#  push:
-#    branches:
-#      - master
-
-jobs:
-  build_dependencies:
-    runs-on: windows-latest
-    steps:
-        - uses: actions/checkout@v3
-        - uses: actions/setup-python@v4
-          with:
-            python-version: '3.10.9'
-
-        - shell: bash
-          run: |
-            python -m pip wheel --no-cache-dir torch torchvision torchaudio ${{ inputs.xformers }} --extra-index-url https://download.pytorch.org/whl/cu118 -r requirements.txt pygit2 -w ./temp_wheel_dir
-            python -m pip install --no-cache-dir ./temp_wheel_dir/*
-            echo installed basic
-            ls -lah temp_wheel_dir
-            mv temp_wheel_dir cu118_python_deps
-            tar cf cu118_python_deps.tar cu118_python_deps
-
-        - uses: actions/cache/save@v3
-          with:
-            path: cu118_python_deps.tar
-            key: ${{ runner.os }}-build-cu118

.github/workflows/windows_release_cu118_package.yml

@@ -1,79 +0,0 @@
-name: "Windows Release cu118 packaging"
-
-on:
-  workflow_dispatch:
-#  push:
-#    branches:
-#      - master
-
-jobs:
-  package_comfyui:
-    permissions:
-        contents: "write"
-        packages: "write"
-        pull-requests: "read"
-    runs-on: windows-latest
-    steps:
-        - uses: actions/cache/restore@v3
-          id: cache
-          with:
-            path: cu118_python_deps.tar
-            key: ${{ runner.os }}-build-cu118
-        - shell: bash
-          run: |
-            mv cu118_python_deps.tar ../
-            cd ..
-            tar xf cu118_python_deps.tar
-            pwd
-            ls
-
-        - uses: actions/checkout@v3
-          with:
-            fetch-depth: 0
-            persist-credentials: false
-        - shell: bash
-          run: |
-            cd ..
-            cp -r ComfyUI ComfyUI_copy
-            curl https://www.python.org/ftp/python/3.10.9/python-3.10.9-embed-amd64.zip -o python_embeded.zip
-            unzip python_embeded.zip -d python_embeded
-            cd python_embeded
-            echo 'import site' >> ./python310._pth
-            curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py
-            ./python.exe get-pip.py
-            ./python.exe -s -m pip install ../cu118_python_deps/*
-            sed -i '1i../ComfyUI' ./python310._pth
-            cd ..
-
-            git clone https://github.com/comfyanonymous/taesd
-            cp taesd/*.pth ./ComfyUI_copy/models/vae_approx/
-
-            mkdir ComfyUI_windows_portable
-            mv python_embeded ComfyUI_windows_portable
-            mv ComfyUI_copy ComfyUI_windows_portable/ComfyUI
-
-            cd ComfyUI_windows_portable
-
-            mkdir update
-            cp -r ComfyUI/.ci/update_windows/* ./update/
-            cp -r ComfyUI/.ci/update_windows_cu118/* ./update/
-            cp -r ComfyUI/.ci/windows_base_files/* ./
-
-            cd ..
-
-            "C:\Program Files\7-Zip\7z.exe" a -t7z -m0=lzma -mx=8 -mfb=64 -md=32m -ms=on -mf=BCJ2 ComfyUI_windows_portable.7z ComfyUI_windows_portable
-            mv ComfyUI_windows_portable.7z ComfyUI/new_ComfyUI_windows_portable_nvidia_cu118_or_cpu.7z
-
-            cd ComfyUI_windows_portable
-            python_embeded/python.exe -s ComfyUI/main.py --quick-test-for-ci --cpu
-
-            ls
-
-        - name: Upload binaries to release
-          uses: svenstaro/upload-release-action@v2
-          with:
-                repo_token: ${{ secrets.GITHUB_TOKEN }}
-                file: new_ComfyUI_windows_portable_nvidia_cu118_or_cpu.7z
-                tag: "latest"
-                overwrite: true
-

.github/workflows/windows_release_dependencies.yml

@@ -41,10 +41,9 @@ jobs:
         - shell: bash
           run: |
             echo "@echo off
-            ..\python_embeded\python.exe .\update.py ..\ComfyUI\\
+            call update_comfyui.bat nopause
             echo -
-            echo This will try to update pytorch and all python dependencies, if you get an error wait for pytorch/xformers to fix their stuff
-            echo You should not be running this anyways unless you really have to
+            echo This will try to update pytorch and all python dependencies.
             echo -
             echo If you just want to update normally, close this and run update_comfyui.bat instead.
             echo -

.github/workflows/windows_release_nightly_pytorch.yml

@@ -68,7 +68,7 @@ jobs:
             cp -r ComfyUI/.ci/update_windows/* ./update/
             cp -r ComfyUI/.ci/windows_base_files/* ./
 
-            echo "..\python_embeded\python.exe .\update.py ..\ComfyUI\\
+            echo "call update_comfyui.bat nopause
             ..\python_embeded\python.exe -s -m pip install --upgrade --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/cu${{ inputs.cu }} -r ../ComfyUI/requirements.txt pygit2
             pause" > ./update/update_comfyui_and_python_dependencies.bat
             cd ..

comfy/clip_model.py

@@ -119,6 +119,9 @@ class CLIPTextModel(torch.nn.Module):
         super().__init__()
         self.num_layers = config_dict["num_hidden_layers"]
         self.text_model = CLIPTextModel_(config_dict, dtype, device, operations)
+        embed_dim = config_dict["hidden_size"]
+        self.text_projection = operations.Linear(embed_dim, embed_dim, bias=False, dtype=dtype, device=device)
+        self.text_projection.weight.copy_(torch.eye(embed_dim))
         self.dtype = dtype
 
     def get_input_embeddings(self):
@@ -128,7 +131,10 @@ class CLIPTextModel(torch.nn.Module):
         self.text_model.embeddings.token_embedding = embeddings
 
     def forward(self, *args, **kwargs):
-        return self.text_model(*args, **kwargs)
+        x = self.text_model(*args, **kwargs)
+        out = self.text_projection(x[2])
+        return (x[0], x[1], out, x[2])
+
 
 class CLIPVisionEmbeddings(torch.nn.Module):
     def __init__(self, embed_dim, num_channels=3, patch_size=14, image_size=224, dtype=None, device=None, operations=None):

comfy/cmd/server.py

@@ -712,11 +712,11 @@ class PromptServer(ExecutorToClientProgress):
 
         for name, dir in nodes.EXTENSION_WEB_DIRS.items():
             self.app.add_routes([
-                web.static('/extensions/' + quote(name), dir, follow_symlinks=True),
+                web.static('/extensions/' + quote(name), dir),
             ])
 
         self.app.add_routes([
-            web.static('/', self.web_root, follow_symlinks=True),
+            web.static('/', self.web_root),
         ])
 
     def get_queue_info(self):

comfy/extra_samplers/uni_pc.py

@@ -358,9 +358,6 @@ class UniPC:
         thresholding=False,
         max_val=1.,
         variant='bh1',
-        noise_mask=None,
-        masked_image=None,
-        noise=None,
     ):
         """Construct a UniPC. 
 
@@ -372,9 +369,6 @@ class UniPC:
         self.predict_x0 = predict_x0
         self.thresholding = thresholding
         self.max_val = max_val
-        self.noise_mask = noise_mask
-        self.masked_image = masked_image
-        self.noise = noise
 
     def dynamic_thresholding_fn(self, x0, t=None):
         """
@@ -391,10 +385,7 @@ class UniPC:
         """
         Return the noise prediction model.
         """
-        if self.noise_mask is not None:
-            return self.model(x, t) * self.noise_mask
-        else:
-            return self.model(x, t)
+        return self.model(x, t)
 
     def data_prediction_fn(self, x, t):
         """
@@ -409,8 +400,6 @@ class UniPC:
             s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1)
             s = expand_dims(torch.maximum(s, self.max_val * torch.ones_like(s).to(s.device)), dims)
             x0 = torch.clamp(x0, -s, s) / s
-        if self.noise_mask is not None:
-            x0 = x0 * self.noise_mask + (1. - self.noise_mask) * self.masked_image
         return x0
 
     def model_fn(self, x, t):
@@ -723,8 +712,6 @@ class UniPC:
             assert timesteps.shape[0] - 1 == steps
             # with torch.no_grad():
             for step_index in trange(steps, disable=disable_pbar):
-                if self.noise_mask is not None:
-                    x = x * self.noise_mask + (1. - self.noise_mask) * (self.masked_image * self.noise_schedule.marginal_alpha(timesteps[step_index]) + self.noise * self.noise_schedule.marginal_std(timesteps[step_index]))
                 if step_index == 0:
                     vec_t = timesteps[0].expand((x.shape[0]))
                     model_prev_list = [self.model_fn(x, vec_t)]
@@ -766,7 +753,7 @@ class UniPC:
                                 model_x = self.model_fn(x, vec_t)
                             model_prev_list[-1] = model_x
                 if callback is not None:
-                    callback(step_index, model_prev_list[-1], x, steps)
+                    callback({'x': x, 'i': step_index, 'denoised': model_prev_list[-1]})
         else:
             raise NotImplementedError()
         # if denoise_to_zero:
@@ -858,7 +845,7 @@ def predict_eps_sigma(model, input, sigma_in, **kwargs):
     return  (input - model(input, sigma_in, **kwargs)) / sigma
 
 
-def sample_unipc(model, noise, image, sigmas, max_denoise, extra_args=None, callback=None, disable=False, noise_mask=None, variant='bh1'):
+def sample_unipc(model, noise, sigmas, extra_args=None, callback=None, disable=False, variant='bh1'):
         timesteps = sigmas.clone()
         if sigmas[-1] == 0:
             timesteps = sigmas[:]
@@ -867,16 +854,7 @@ def sample_unipc(model, noise, image, sigmas, max_denoise, extra_args=None, call
             timesteps = sigmas.clone()
         ns = SigmaConvert()
 
-        if image is not None:
-            img = image * ns.marginal_alpha(timesteps[0])
-            if max_denoise:
-                noise_mult = 1.0
-            else:
-                noise_mult = ns.marginal_std(timesteps[0])
-            img += noise * noise_mult
-        else:
-            img = noise
-
+        noise = noise / torch.sqrt(1.0 + timesteps[0] ** 2.0)
         model_type = "noise"
 
         model_fn = model_wrapper(
@@ -888,7 +866,10 @@ def sample_unipc(model, noise, image, sigmas, max_denoise, extra_args=None, call
         )
 
         order = min(3, len(timesteps) - 2)
-        uni_pc = UniPC(model_fn, ns, predict_x0=True, thresholding=False, noise_mask=noise_mask, masked_image=image, noise=noise, variant=variant)
-        x = uni_pc.sample(img, timesteps=timesteps, skip_type="time_uniform", method="multistep", order=order, lower_order_final=True, callback=callback, disable_pbar=disable)
+        uni_pc = UniPC(model_fn, ns, predict_x0=True, thresholding=False, variant=variant)
+        x = uni_pc.sample(noise, timesteps=timesteps, skip_type="time_uniform", method="multistep", order=order, lower_order_final=True, callback=callback, disable_pbar=disable)
         x /= ns.marginal_alpha(timesteps[-1])
         return x
+
+def sample_unipc_bh2(model, noise, sigmas, extra_args=None, callback=None, disable=False):
+    return sample_unipc(model, noise, sigmas, extra_args, callback, disable, variant='bh2')

comfy/lora.py

@@ -197,6 +197,15 @@ def model_lora_keys_clip(model, key_map={}):
                     key_map[lora_key] = k
                     lora_key = "text_encoder.text_model.encoder.layers.{}.{}".format(b, c) #diffusers lora
                     key_map[lora_key] = k
+                    lora_key = "lora_prior_te_text_model_encoder_layers_{}_{}".format(b, LORA_CLIP_MAP[c]) #cascade lora: TODO put lora key prefix in the model config
+                    key_map[lora_key] = k
+
+
+    k = "clip_g.transformer.text_projection.weight"
+    if k in sdk:
+        key_map["lora_prior_te_text_projection"] = k #cascade lora?
+        # key_map["text_encoder.text_projection"] = k #TODO: check if other lora have the text_projection too
+        # key_map["lora_te_text_projection"] = k
 
     return key_map
 
@@ -207,6 +216,7 @@ def model_lora_keys_unet(model, key_map={}):
         if k.startswith("diffusion_model.") and k.endswith(".weight"):
             key_lora = k[len("diffusion_model."):-len(".weight")].replace(".", "_")
             key_map["lora_unet_{}".format(key_lora)] = k
+            key_map["lora_prior_unet_{}".format(key_lora)] = k #cascade lora: TODO put lora key prefix in the model config
 
     diffusers_keys = utils.unet_to_diffusers(model.model_config.unet_config)
     for k in diffusers_keys:

comfy/nodes/base_nodes.py

@@ -987,7 +987,7 @@ class GLIGENTextBoxApply:
 
     def append(self, conditioning_to, clip, gligen_textbox_model, text, width, height, x, y):
         c = []
-        cond, cond_pooled = clip.encode_from_tokens(clip.tokenize(text), return_pooled=True)
+        cond, cond_pooled = clip.encode_from_tokens(clip.tokenize(text), return_pooled="unprojected")
         for t in conditioning_to:
             n = [t[0], t[1].copy()]
             position_params = [(cond_pooled, height // 8, width // 8, y // 8, x // 8)]

comfy/samplers.py

@@ -211,6 +211,7 @@ def calc_cond_uncond_batch(model, cond, uncond, x_in, timestep, model_options):
                         cur_patches[p] = cur_patches[p] + patches[p]
                     else:
                         cur_patches[p] = patches[p]
+                transformer_options["patches"] = cur_patches
             else:
                 transformer_options["patches"] = patches
 
@@ -516,14 +517,9 @@ class Sampler:
         sigma = float(sigmas[0])
         return math.isclose(max_sigma, sigma, rel_tol=1e-05) or sigma > max_sigma
 
-class UNIPC(Sampler):
-    def sample(self, model_wrap, sigmas, extra_args, callback, noise, latent_image=None, denoise_mask=None, disable_pbar=False):
-        return uni_pc.sample_unipc(model_wrap, noise, latent_image, sigmas, max_denoise=self.max_denoise(model_wrap, sigmas), extra_args=extra_args, noise_mask=denoise_mask, callback=callback, disable=disable_pbar)
-
-class UNIPCBH2(Sampler):
-    def sample(self, model_wrap, sigmas, extra_args, callback, noise, latent_image=None, denoise_mask=None, disable_pbar=False):
-        return uni_pc.sample_unipc(model_wrap, noise, latent_image, sigmas, max_denoise=self.max_denoise(model_wrap, sigmas), extra_args=extra_args, noise_mask=denoise_mask, callback=callback, variant='bh2', disable=disable_pbar)
-
+KSAMPLER_NAMES = ["euler", "euler_ancestral", "heun", "heunpp2","dpm_2", "dpm_2_ancestral",
+                  "lms", "dpm_fast", "dpm_adaptive", "dpmpp_2s_ancestral", "dpmpp_sde", "dpmpp_sde_gpu",
+                  "dpmpp_2m", "dpmpp_2m_sde", "dpmpp_2m_sde_gpu", "dpmpp_3m_sde", "dpmpp_3m_sde_gpu", "ddpm", "lcm"]
 
 class KSAMPLER(Sampler):
     def __init__(self, sampler_function, extra_options={}, inpaint_options={}):
@@ -638,9 +634,9 @@ def calculate_sigmas_scheduler(model, scheduler_name, steps):
 
 def sampler_object(name):
     if name == "uni_pc":
-        sampler = UNIPC()
+        sampler = KSAMPLER(uni_pc.sample_unipc)
     elif name == "uni_pc_bh2":
-        sampler = UNIPCBH2()
+        sampler = KSAMPLER(uni_pc.sample_unipc_bh2)
     elif name == "ddim":
         sampler = ksampler("euler", inpaint_options={"random": True})
     else:

comfy/sd.py

@@ -52,7 +52,7 @@ def load_clip_weights(model, sd):
         if ids.dtype == torch.float32:
             sd['cond_stage_model.transformer.text_model.embeddings.position_ids'] = ids.round()
 
-    sd = utils.transformers_convert(sd, "cond_stage_model.model.", "cond_stage_model.transformer.text_model.", 24)
+    sd = utils.clip_text_transformers_convert(sd, "cond_stage_model.model.", "cond_stage_model.transformer.")
     return load_model_weights(model, sd)
 
 
@@ -123,10 +123,13 @@ class CLIP:
         return self.tokenizer.tokenize_with_weights(text, return_word_ids)
 
     def encode_from_tokens(self, tokens, return_pooled=False):
+        self.cond_stage_model.reset_clip_options()
+
         if self.layer_idx is not None:
-            self.cond_stage_model.clip_layer(self.layer_idx)
-        else:
-            self.cond_stage_model.reset_clip_layer()
+            self.cond_stage_model.set_clip_options({"layer": self.layer_idx})
+
+        if return_pooled == "unprojected":
+            self.cond_stage_model.set_clip_options({"projected_pooled": False})
 
         self.load_model()
         cond, pooled = self.cond_stage_model.encode_token_weights(tokens)
@@ -361,7 +364,10 @@ def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DI
 
     for i in range(len(clip_data)):
         if "transformer.resblocks.0.ln_1.weight" in clip_data[i]:
-            clip_data[i] = utils.transformers_convert(clip_data[i], "", "text_model.", 32)
+            clip_data[i] = utils.clip_text_transformers_convert(clip_data[i], "", "")
+        else:
+            if "text_projection" in clip_data[i]:
+                clip_data[i]["text_projection.weight"] = clip_data[i]["text_projection"].transpose(0, 1) #old models saved with the CLIPSave node
 
     clip_target = EmptyClass()
     clip_target.params = {}

comfy/sd1_clip.py

@@ -68,7 +68,7 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
     ]
     def __init__(self, version="openai/clip-vit-large-patch14", device="cpu", max_length=77,
                  freeze=True, layer="last", layer_idx=None, textmodel_json_config=None, dtype=None, model_class=clip_model.CLIPTextModel,
-                 special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=True, enable_attention_masks=False):  # clip-vit-base-patch32
+                 special_tokens={"start": 49406, "end": 49407, "pad": 49407}, layer_norm_hidden_state=True, enable_attention_masks=False, return_projected_pooled=True):  # clip-vit-base-patch32
         super().__init__()
         assert layer in self.LAYERS
 
@@ -89,16 +89,18 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         self.layer = layer
         self.layer_idx = None
         self.special_tokens = special_tokens
-        self.text_projection = torch.nn.Parameter(torch.eye(self.transformer.get_input_embeddings().weight.shape[1]))
+
         self.logit_scale = torch.nn.Parameter(torch.tensor(4.6055))
         self.enable_attention_masks = enable_attention_masks
 
         self.layer_norm_hidden_state = layer_norm_hidden_state
+        self.return_projected_pooled = return_projected_pooled
+
         if layer == "hidden":
             assert layer_idx is not None
             assert abs(layer_idx) < self.num_layers
-            self.clip_layer(layer_idx)
-        self.layer_default = (self.layer, self.layer_idx)
+            self.set_clip_options({"layer": layer_idx})
+        self.options_default = (self.layer, self.layer_idx, self.return_projected_pooled)
 
     def freeze(self):
         self.transformer = self.transformer.eval()
@@ -106,16 +108,19 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         for param in self.parameters():
             param.requires_grad = False
 
-    def clip_layer(self, layer_idx):
-        if abs(layer_idx) > self.num_layers:
+    def set_clip_options(self, options):
+        layer_idx = options.get("layer", self.layer_idx)
+        self.return_projected_pooled = options.get("projected_pooled", self.return_projected_pooled)
+        if layer_idx is None or abs(layer_idx) > self.num_layers:
             self.layer = "last"
         else:
             self.layer = "hidden"
             self.layer_idx = layer_idx
 
-    def reset_clip_layer(self):
-        self.layer = self.layer_default[0]
-        self.layer_idx = self.layer_default[1]
+    def reset_clip_options(self):
+        self.layer = self.options_default[0]
+        self.layer_idx = self.options_default[1]
+        self.return_projected_pooled = self.options_default[2]
 
     def set_up_textual_embeddings(self, tokens, current_embeds):
         out_tokens = []
@@ -180,23 +185,19 @@ class SDClipModel(torch.nn.Module, ClipTokenWeightEncoder):
         else:
             z = outputs[1]
 
-        if outputs[2] is not None:
-            pooled_output = outputs[2].float()
-        else:
-            pooled_output = None
+        pooled_output = None
+        if len(outputs) >= 3:
+            if not self.return_projected_pooled and len(outputs) >= 4 and outputs[3] is not None:
+                pooled_output = outputs[3].float()
+            elif outputs[2] is not None:
+                pooled_output = outputs[2].float()
 
-        if self.text_projection is not None and pooled_output is not None:
-            pooled_output = pooled_output.float().to(self.text_projection.device) @ self.text_projection.float()
         return z.float(), pooled_output
 
     def encode(self, tokens):
         return self(tokens)
 
     def load_sd(self, sd):
-        if "text_projection" in sd:
-            self.text_projection[:] = sd.pop("text_projection")
-        if "text_projection.weight" in sd:
-            self.text_projection[:] = sd.pop("text_projection.weight").transpose(0, 1)
         return self.transformer.load_state_dict(sd, strict=False)
 
 def parse_parentheses(string):
@@ -515,11 +516,11 @@ class SD1ClipModel(torch.nn.Module):
         self.clip = "clip_{}".format(self.clip_name)
         setattr(self, self.clip, clip_model(device=device, dtype=dtype, **kwargs))
 
-    def clip_layer(self, layer_idx):
-        getattr(self, self.clip).clip_layer(layer_idx)
+    def set_clip_options(self, options):
+        getattr(self, self.clip).set_clip_options(options)
 
-    def reset_clip_layer(self):
-        getattr(self, self.clip).reset_clip_layer()
+    def reset_clip_options(self):
+        getattr(self, self.clip).reset_clip_options()
 
     def encode_token_weights(self, token_weight_pairs):
         token_weight_pairs = token_weight_pairs[self.clip_name]

comfy/sdxl_clip.py

@@ -40,13 +40,13 @@ class SDXLClipModel(torch.nn.Module):
         self.clip_l = sd1_clip.SDClipModel(layer="hidden", layer_idx=-2, device=device, dtype=dtype, layer_norm_hidden_state=False)
         self.clip_g = SDXLClipG(device=device, dtype=dtype)
 
-    def clip_layer(self, layer_idx):
-        self.clip_l.clip_layer(layer_idx)
-        self.clip_g.clip_layer(layer_idx)
+    def set_clip_options(self, options):
+        self.clip_l.set_clip_options(options)
+        self.clip_g.set_clip_options(options)
 
-    def reset_clip_layer(self):
-        self.clip_g.reset_clip_layer()
-        self.clip_l.reset_clip_layer()
+    def reset_clip_options(self):
+        self.clip_g.reset_clip_options()
+        self.clip_l.reset_clip_options()
 
     def encode_token_weights(self, token_weight_pairs):
         token_weight_pairs_g = token_weight_pairs["g"]

comfy/supported_models.py

@@ -75,7 +75,7 @@ class SD20(supported_models_base.BASE):
         replace_prefix["conditioner.embedders.0.model."] = "clip_h." #SD2 in sgm format
         replace_prefix["cond_stage_model.model."] = "clip_h."
         state_dict = utils.state_dict_prefix_replace(state_dict, replace_prefix, filter_keys=True)
-        state_dict = utils.transformers_convert(state_dict, "clip_h.", "clip_h.transformer.text_model.", 24)
+        state_dict = utils.clip_text_transformers_convert(state_dict, "clip_h.", "clip_h.transformer.")
         return state_dict
 
     def process_clip_state_dict_for_saving(self, state_dict):
@@ -134,7 +134,7 @@ class SDXLRefiner(supported_models_base.BASE):
         replace_prefix["conditioner.embedders.0.model."] = "clip_g."
         state_dict = utils.state_dict_prefix_replace(state_dict, replace_prefix, filter_keys=True)
 
-        state_dict = utils.transformers_convert(state_dict, "clip_g.", "clip_g.transformer.text_model.", 32)
+        state_dict = utils.clip_text_transformers_convert(state_dict, "clip_g.", "clip_g.transformer.")
         state_dict = utils.state_dict_key_replace(state_dict, keys_to_replace)
         return state_dict
 
@@ -182,10 +182,8 @@ class SDXL(supported_models_base.BASE):
         replace_prefix["conditioner.embedders.1.model."] = "clip_g."
         state_dict = utils.state_dict_prefix_replace(state_dict, replace_prefix, filter_keys=True)
 
-        state_dict = utils.transformers_convert(state_dict, "clip_g.", "clip_g.transformer.text_model.", 32)
-        keys_to_replace["clip_g.text_projection.weight"] = "clip_g.text_projection"
-
         state_dict = utils.state_dict_key_replace(state_dict, keys_to_replace)
+        state_dict = utils.clip_text_transformers_convert(state_dict, "clip_g.", "clip_g.transformer.")
         return state_dict
 
     def process_clip_state_dict_for_saving(self, state_dict):
@@ -338,6 +336,12 @@ class Stable_Cascade_C(supported_models_base.BASE):
                     state_dict[k_to] = weights[shape_from*x:shape_from*(x + 1)]
         return state_dict
 
+    def process_clip_state_dict(self, state_dict):
+        state_dict = utils.state_dict_prefix_replace(state_dict, {k: "" for k in self.text_encoder_key_prefix}, filter_keys=True)
+        if "clip_g.text_projection" in state_dict:
+            state_dict["clip_g.transformer.text_projection.weight"] = state_dict.pop("clip_g.text_projection").transpose(0, 1)
+        return state_dict
+
     def get_model(self, state_dict, prefix="", device=None):
         out = model_base.StableCascade_C(self, device=device)
         return out

comfy/utils.py

@@ -100,8 +100,22 @@ def transformers_convert(sd, prefix_from, prefix_to, number):
                     p = ["self_attn.q_proj", "self_attn.k_proj", "self_attn.v_proj"]
                     k_to = "{}encoder.layers.{}.{}.{}".format(prefix_to, resblock, p[x], y)
                     sd[k_to] = weights[shape_from*x:shape_from*(x + 1)]
+
     return sd
 
+def clip_text_transformers_convert(sd, prefix_from, prefix_to):
+    sd = transformers_convert(sd, prefix_from, "{}text_model.".format(prefix_to), 32)
+
+    tp = "{}text_projection.weight".format(prefix_from)
+    if tp in sd:
+        sd["{}text_projection.weight".format(prefix_to)] = sd.pop(tp)
+
+    tp = "{}text_projection".format(prefix_from)
+    if tp in sd:
+        sd["{}text_projection.weight".format(prefix_to)] = sd.pop(tp).transpose(0, 1)
+    return sd
+
+
 UNET_MAP_ATTENTIONS = {
     "proj_in.weight",
     "proj_in.bias",

comfy_extras/nodes/nodes_perpneg.py

@@ -33,7 +33,7 @@ class PerpNeg:
 
             pos = noise_pred_pos - noise_pred_nocond
             neg = noise_pred_neg - noise_pred_nocond
-            perp = ((torch.mul(pos, neg).sum())/(torch.norm(neg)**2)) * neg
+            perp = neg - ((torch.mul(neg, pos).sum())/(torch.norm(pos)**2)) * pos
             perp_neg = perp * neg_scale
             cfg_result = noise_pred_nocond + cond_scale*(pos - perp_neg)
             cfg_result = x - cfg_result

custom_nodes/example_node.py.example

@@ -103,6 +103,9 @@ class Example:
     #def IS_CHANGED(s, image, string_field, int_field, float_field, print_to_screen):
     #    return ""
 
+# Set the web directory, any .js file in that directory will be loaded by the frontend as a frontend extension
+# WEB_DIRECTORY = "./somejs"
+
 # A dictionary that contains all nodes you want to export with their names
 # NOTE: names should be globally unique
 NODE_CLASS_MAPPINGS = {