Release v0.4.0

.github/workflows/package-release.yml

@@ -6,8 +6,25 @@ on:
   workflow_dispatch:
 
 jobs:
-  windows-cuda:
-    runs-on: windows-latest
+  package-release:
+    strategy:
+      matrix:
+        platform:
+          - requirements: win-linux-cuda.txt
+            os: windows-latest
+            filename: windows-cuda
+          - requirements: win-dml.txt
+            os: windows-latest
+            filename:  windows-directml
+          - requirements: mac-mps-cpu.txt
+            os: macos-14
+            filename: macos-arm
+        version:
+          - python: '3.10'
+            filename_suffix: ''
+          - python: '3.11'
+            filename_suffix: '-4-1'
+    runs-on: ${{ matrix.platform.os }}
     steps:
       - name: Checkout repository
         uses: actions/checkout@v3
@@ -16,12 +33,12 @@ jobs:
       - name: Setup Python
         uses: actions/setup-python@v4
         with:
-          python-version: '3.10'
+          python-version: ${{ matrix.version.python }}
           cache: 'pip'
-          cache-dependency-path: '**/win-linux-cuda.txt'
+          cache-dependency-path: '**/${{ matrix.platform.requirements }}'
       - name: Install dependencies into target
         shell: bash
-        run: 'python -m pip install -r requirements/win-linux-cuda.txt --no-cache-dir --target .python_dependencies'
+        run: 'python -m pip install -r requirements/${{ matrix.platform.requirements }} --no-cache-dir --target .python_dependencies'
         working-directory: dream_textures
       - name: Zip dependencies with long paths
         shell: bash
@@ -30,72 +47,10 @@ jobs:
         uses: thedoctor0/zip-release@main
         with:
           type: zip
-          filename: dream_textures-windows-cuda.zip
+          filename: dream_textures-${{ matrix.platform.filename }}${{ matrix.version.filename_suffix }}.zip
           exclusions: '*.git*'
       - name: Archive and upload artifact
         uses: actions/upload-artifact@v3
         with:
-          name: dream_textures-windows-cuda
-          path: dream_textures-windows-cuda.zip
-  windows-directml:
-    runs-on: windows-latest
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v3
-        with:
-          path: dream_textures
-      - name: Setup Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: '3.10'
-          cache: 'pip'
-          cache-dependency-path: '**/win-dml.txt'
-      - name: Install dependencies into target
-        shell: bash
-        run: 'python -m pip install -r requirements/win-dml.txt --no-cache-dir --target .python_dependencies'
-        working-directory: dream_textures
-      - name: Zip dependencies with long paths
-        shell: bash
-        run: 'python ./dream_textures/scripts/zip_dependencies.py'
-      - name: Archive Release
-        uses: thedoctor0/zip-release@main
-        with:
-          type: zip
-          filename: dream_textures-windows-directml.zip
-          exclusions: '*.git*'
-      - name: Archive and upload artifact
-        uses: actions/upload-artifact@v3
-        with:
-          name: dream_textures-windows-directml
-          path: dream_textures-windows-directml.zip
-  macos-arm:
-    runs-on: macos-14
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v3
-        with:
-          path: dream_textures
-      - name: Setup Python
-        uses: actions/setup-python@v4
-        with:
-          python-version: '3.10'
-          cache: 'pip'
-          cache-dependency-path: '**/mac-mps-cpu.txt'
-      - name: Install dependencies into target
-        shell: bash
-        run: 'python -m pip install -r requirements/mac-mps-cpu.txt --no-cache-dir --target .python_dependencies'
-        working-directory: dream_textures
-      - name: Zip dependencies with long paths
-        shell: bash
-        run: 'python ./dream_textures/scripts/zip_dependencies.py'
-      - name: Archive Release
-        uses: thedoctor0/zip-release@main
-        with:
-          type: zip
-          filename: dream_textures-macos-arm.zip
-          exclusions: '*.git*'
-      - name: Archive and upload artifact
-        uses: actions/upload-artifact@v3
-        with:
-          name: dream_textures-macos-arm
-          path: dream_textures-macos-arm.zip
+          name: dream_textures-${{ matrix.platform.filename }}${{ matrix.version.filename_suffix }}
+          path: dream_textures-${{ matrix.platform.filename }}${{ matrix.version.filename_suffix }}.zip

__init__.py

@@ -16,7 +16,7 @@
     "author": "Dream Textures contributors",
     "description": "Use Stable Diffusion to generate unique textures straight from the shader editor.",
     "blender": (3, 1, 0),
-    "version": (0, 3, 1),
+    "version": (0, 4, 0),
     "location": "Image Editor -> Sidebar -> Dream",
     "category": "Paint"
 }

engine/engine.py

@@ -87,7 +87,7 @@ def execute(self, context):
         node_out = node_tree.nodes.new(type="NodeGroupOutput")
         # Blender 4.0 uses a new API for in- and outputs
         if bpy.app.version[0] > 3:
-            node_tree.interface.new_socket('Image', description="Output of the final image.", in_out='OUTPUT', type='NodeSocketColor')
+            node_tree.interface.new_socket('Image', description="Output of the final image.", in_out='OUTPUT', socket_type='NodeSocketColor')
         else:
             node_tree.outputs.new('NodeSocketColor','Image')
         node_out.location = (400, 200)

generator_process/actions/depth_to_image.py

@@ -194,7 +194,7 @@ def __call__(
             width = width or self.unet.config.sample_size * self.vae_scale_factor
 
             # 1. Check inputs
-            self.check_inputs(prompt, height, width, strength, callback_steps)
+            self.check_inputs(prompt=prompt, image=image, mask_image=depth_image, height=height, width=width, strength=strength, callback_steps=callback_steps, output_type=output_type)
 
             # 2. Define call parameters
             batch_size = 1 if isinstance(prompt, str) else len(prompt)
@@ -366,11 +366,10 @@ def __call__(
 
     # Inference
     with torch.inference_mode() if device not in ('mps', "dml") else nullcontext():
-        def callback(pipe, step, timestep, callback_kwargs):
+        def callback(step, _, latents):
             if future.check_cancelled():
                 raise InterruptedError()
-            future.add_response(step_latents(pipe, step_preview_mode, callback_kwargs["latents"], generator, step, steps))
-            return callback_kwargs
+            future.add_response(step_latents(pipe, step_preview_mode, latents, generator, step, steps))
         try:
             result = pipe(
                 prompt=prompt,
@@ -383,7 +382,7 @@ def callback(pipe, step, timestep, callback_kwargs):
                 num_inference_steps=steps,
                 guidance_scale=cfg_scale,
                 generator=generator,
-                callback_on_step_end=callback,
+                callback=callback,
                 callback_steps=1,
                 output_type="np"
             )

generator_process/actions/huggingface_hub.py

@@ -163,6 +163,7 @@ def hf_snapshot_download(
 ):
     from huggingface_hub import snapshot_download, repo_info
     from diffusers import StableDiffusionPipeline
+    from diffusers.pipelines.pipeline_utils import variant_compatible_siblings
 
     future = Future()
     yield future
@@ -172,10 +173,12 @@ def hf_snapshot_download(
     files = [file.rfilename for file in info.siblings]
 
     if "model_index.json" in files:
+        # check if the variant files are available before trying to download them
+        _, variant_files = variant_compatible_siblings(files, variant=variant)
         StableDiffusionPipeline.download(
             model,
             use_auth_token=token,
-            variant=variant,
+            variant=variant if len(variant_files) > 0 else None,
             resume_download=resume_download,
         )
     elif "config.json" in files:

generator_process/actor.py

@@ -33,6 +33,14 @@ def _load_dependencies():
         python3_path = os.path.abspath(os.path.join(sys.executable, "..\\..\\..\\..\\python3.dll"))
         if os.path.exists(python3_path):
             os.add_dll_directory(os.path.dirname(python3_path))
+
+        # fix for OSError: [WinError 126] The specified module could not be found. Error loading "...\dream_textures\.python_dependencies\torch\lib\shm.dll" or one of its dependencies.
+        # Allows for shm.dll from torch==2.3.0 to access dependencies from mkl==2021.4.0
+        # These DLL dependencies are not in the usual places that torch would look at due to being pip installed to a target directory.
+        mkl_bin = absolute_path(".python_dependencies\\Library\\bin")
+        if os.path.exists(mkl_bin):
+            os.add_dll_directory(mkl_bin)
+
     if os.path.exists(absolute_path(".python_dependencies.zip")):
         sys.path.insert(1, absolute_path(".python_dependencies.zip"))
         _patch_zip_direct_transformers_import()

generator_process/directml_patches.py

@@ -7,16 +7,6 @@
 active_dml_patches: list | None = None
 
 
-def baddbmm(input, batch1, batch2, *, beta=1, alpha=1, out=None, pre_patch):
-    if input.device.type == "dml" and beta == 0:
-        if out is not None:
-            torch.bmm(batch1, batch2, out=out)
-            out *= alpha
-            return out
-        return alpha * (batch1 @ batch2)
-    return pre_patch(input, batch1, batch2, beta=beta, alpha=alpha, out=out)
-
-
 def pad(input, pad, mode="constant", value=None, *, pre_patch):
     if input.device.type == "dml" and mode == "constant":
         pad_dims = torch.tensor(pad, dtype=torch.int32).view(-1, 2).flip(0)
@@ -39,10 +29,10 @@ def pad(input, pad, mode="constant", value=None, *, pre_patch):
     return pre_patch(input, pad, mode=mode, value=value)
 
 
-def getitem(self, key, *, pre_patch):
-    if isinstance(key, Tensor) and "dml" in [self.device.type, key.device.type] and key.numel() == 1:
-        return pre_patch(self, int(key))
-    return pre_patch(self, key)
+def layer_norm(input, normalized_shape, weight = None, bias = None, eps = 1e-05, *, pre_patch):
+    if input.device.type == "dml":
+        return pre_patch(input.contiguous(), normalized_shape, weight, bias, eps)
+    return pre_patch(input, normalized_shape, weight, bias, eps)
 
 
 def retry_OOM(module):
@@ -110,17 +100,9 @@ def dml_patch_method(object, name, patched):
         setattr(object, name, functools.partialmethod(patched, pre_patch=original))
         active_dml_patches.append({"object": object, "name": name, "original": original})
 
-    # Not all places where the patches have an effect are necessarily listed.
-
-    # diffusers.models.attention_processor.Attention.get_attention_scores()
-    # diffusers.models.attention.AttentionBlock.forward()
-    # Diffusers implementation gives torch.empty() tensors with beta=0 to baddbmm(), which may contain NaNs.
-    # DML implementation doesn't properly ignore input argument with beta=0 and causes NaN propagation.
-    dml_patch(torch, "baddbmm", baddbmm)
-
     dml_patch(torch.nn.functional, "pad", pad)
-    # DDIMScheduler.step(), PNDMScheduler.step(), No error messages or crashes, just may randomly freeze.
-    dml_patch_method(Tensor, "__getitem__", getitem)
+
+    dml_patch(torch.nn.functional, "layer_norm", layer_norm)
 
     def decorate_forward(name, module):
         """Helper function to better find which modules DML fails in as it often does

image_utils.py

@@ -689,7 +689,10 @@ def np_to_render_pass(
         array = to_dtype(array, dtype)
     if top_to_bottom:
         array = np.flipud(array)
-    render_pass.rect.foreach_set(array.reshape(-1, render_pass.channels))
+    if BLENDER_VERSION >= (4, 1, 0):
+        render_pass.rect.foreach_set(array.reshape(-1))
+    else:
+        render_pass.rect.foreach_set(array.reshape(-1, render_pass.channels))
 
 
 def _mode(array, mode):

requirements/linux-rocm.txt

@@ -5,8 +5,8 @@ accelerate
 huggingface_hub
 controlnet-aux==0.0.7
 
---extra-index-url https://download.pytorch.org/whl/rocm5.6/
-torch>=2.1
+--extra-index-url https://download.pytorch.org/whl/nightly/rocm6.1/
+torch>=2.3.1
 
 # Original SD checkpoint conversion
 pytorch-lightning
@@ -16,4 +16,4 @@ omegaconf
 scipy # LMSDiscreteScheduler
 
 opencolorio==2.3.2 # color management
-matplotlib
+matplotlib

version.py

@@ -1,4 +1,4 @@
-VERSION = (0, 3, 1)
+VERSION = (0, 4, 0)
 def version_tag(version):
     return f"{version[0]}.{version[1]}.{version[2]}"