Add ControlNet Tile e2e test

scripts/prepare_test_weights.py

@@ -291,6 +291,13 @@ def download_controlnet():
         ]
         download_files(urls, net_folder)
 
+    tile_folder = os.path.join(base_folder, "control_v11f1e_sd15_tile")
+    urls = [
+        "https://huggingface.co/lllyasviel/control_v11f1e_sd15_tile/raw/main/config.json",
+        "https://huggingface.co/lllyasviel/control_v11f1e_sd15_tile/resolve/main/diffusion_pytorch_model.bin",
+    ]
+    download_files(urls, tile_folder)
+
     mfidabel_folder = os.path.join(test_weights_dir, "mfidabel", "controlnet-segment-anything")
     urls = [
         "https://huggingface.co/mfidabel/controlnet-segment-anything/raw/main/config.json",
@@ -597,6 +604,12 @@ def convert_controlnet():
         "tests/weights/controlnet/mfidabel_controlnet-segment-anything.safetensors",
         expected_hash="d536eebb",
     )
+    run_conversion_script(
+        "convert_diffusers_controlnet.py",
+        "tests/weights/lllyasviel/control_v11f1e_sd15_tile",
+        "tests/weights/controlnet/lllyasviel_control_v11f1e_sd15_tile.safetensors",
+        expected_hash="42463af8",
+    )
 
 
 def convert_unclip():

tests/e2e/test_diffusion.py

@@ -194,6 +194,14 @@ def controlnet_data_scale_decay(
     yield (cn_name, condition_image, expected_image, weights_path)
 
 
+@pytest.fixture(scope="module")
+def controlnet_data_tile(ref_path: Path, test_weights_path: Path) -> tuple[Image.Image, Image.Image, Path]:
+    condition_image = _img_open(ref_path / f"low_res_dog.png").convert("RGB").resize((1024, 1024))  # type: ignore
+    expected_image = _img_open(ref_path / f"expected_controlnet_tile.png").convert("RGB")
+    weights_path = test_weights_path / "controlnet" / "lllyasviel_control_v11f1e_sd15_tile.safetensors"
+    return condition_image, expected_image, weights_path
+
+
 @pytest.fixture(scope="module")
 def controlnet_data_canny(ref_path: Path, test_weights_path: Path) -> tuple[str, Image.Image, Image.Image, Path]:
     cn_name = "canny"
@@ -1091,6 +1099,43 @@ def test_diffusion_controlnet(
     ensure_similar_images(predicted_image, expected_image, min_psnr=35, min_ssim=0.98)
 
 
+@no_grad()
+def test_diffusion_controlnet_tile_upscale(
+    sd15_std: StableDiffusion_1,
+    controlnet_data_tile: tuple[Image.Image, Image.Image, Path],
+    test_device: torch.device,
+):
+    sd15 = sd15_std
+
+    condition_image, expected_image, cn_weights_path = controlnet_data_tile
+
+    controlnet: SD1ControlnetAdapter = SD1ControlnetAdapter(
+        sd15.unet, name="tile", scale=1.0, weights=load_from_safetensors(cn_weights_path)
+    ).inject()
+
+    cn_condition = image_to_tensor(condition_image, device=test_device)
+
+    prompt = "best quality"
+    negative_prompt = "blur, lowres, bad anatomy, bad hands, cropped, worst quality"
+    clip_text_embedding = sd15.compute_clip_text_embedding(text=prompt, negative_text=negative_prompt)
+
+    manual_seed(0)
+    x = sd15.init_latents((1024, 1024), condition_image).to(test_device)
+
+    for step in sd15.steps:
+        controlnet.set_controlnet_condition(cn_condition)
+        x = sd15(
+            x,
+            step=step,
+            clip_text_embedding=clip_text_embedding,
+            condition_scale=7.5,
+        )
+    predicted_image = sd15.lda.latents_to_image(x)
+
+    # Note: rather large tolerances are used on purpose here (loose comparison with diffusers' output)
+    ensure_similar_images(predicted_image, expected_image, min_psnr=24, min_ssim=0.75)
+
+
 @no_grad()
 def test_diffusion_controlnet_scale_decay(
     sd15_std: StableDiffusion_1,

tests/e2e/test_diffusion_ref/README.md

@@ -57,8 +57,7 @@ Special cases:
     - `expected_controllora_PyraCanny+CPDS.png`
     - `expected_controllora_disabled.png`
     - `expected_style_aligned.png`
-    - `expected_controlnet_<name>.png` (canny|depth|lineart|normals|sam|stack)
-    - `expected_controlnet_<name>_scale_decay.png` (canny)
+    - `expected_controlnet_canny_scale_decay.png` 
 
 ## Other images
 
@@ -90,6 +89,9 @@ Special cases:
 
 - `cutecat_guide_PyraCanny.png` and `cutecat_guide_CPDS.png` were [generated inside Fooocus](https://github.com/lllyasviel/Fooocus/blob/e8d88d3e250e541c6daf99d6ef734e8dc3cfdc7f/extras/preprocessors.py).
 
+- `low_res_dog.png` and `expected_controlnet_tile.png` are taken from Diffusers [documentation](https://huggingface.co/lllyasviel/control_v11f1e_sd15_tile/tree/main/images), respectively named
+`original.png` and `output.png`.
+
 ## VAE without randomness
 
 ```diff