fix broken self-attention guidance with ip-adapter

The #168 and #177 refactorings caused this regression. A new end-to-end
test has been added for proper coverage.

(This fix will be revisited at some point)

README.md

@@ -156,7 +156,7 @@ print("done: see output.png")
 
 You should get:
 
-![dropy slime output](https://raw.githubusercontent.com/finegrain-ai/refiners/main/assets/dropy_slime_9752.png)
+![dropy slime output](https://raw.githubusercontent.com/finegrain-ai/refiners/main/assets/expected_dropy_slime_9752.png)
 
 ### Training
 

scripts/prepare_test_weights.py

@@ -209,6 +209,16 @@ def download_sdxl(hf_repo_id: str = "stabilityai/stable-diffusion-xl-base-1.0"):
     download_sd_tokenizer(hf_repo_id, "tokenizer_2")
 
 
+def download_vae_fp16_fix():
+    download_files(
+        urls=[
+            "https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/raw/main/config.json",
+            "https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/resolve/main/diffusion_pytorch_model.safetensors",
+        ],
+        dest_folder=os.path.join(test_weights_dir, "madebyollin", "sdxl-vae-fp16-fix"),
+    )
+
+
 def download_vae_ft_mse():
     download_files(
         urls=[
@@ -433,6 +443,17 @@ def convert_vae_ft_mse():
     )
 
 
+def convert_vae_fp16_fix():
+    run_conversion_script(
+        "convert_diffusers_autoencoder_kl.py",
+        "tests/weights/madebyollin/sdxl-vae-fp16-fix",
+        "tests/weights/sdxl-lda-fp16-fix.safetensors",
+        additional_args=["--subfolder", "''"],
+        half=True,
+        expected_hash="98c7e998",
+    )
+
+
 def convert_lora():
     os.makedirs("tests/weights/loras", exist_ok=True)
     run_conversion_script(
@@ -610,6 +631,7 @@ def download_all():
     download_sd15("runwayml/stable-diffusion-inpainting")
     download_sdxl("stabilityai/stable-diffusion-xl-base-1.0")
     download_vae_ft_mse()
+    download_vae_fp16_fix()
     download_lora()
     download_preprocessors()
     download_controlnet()
@@ -624,6 +646,7 @@ def convert_all():
     convert_sd15()
     convert_sdxl()
     convert_vae_ft_mse()
+    convert_vae_fp16_fix()
     convert_lora()
     convert_preprocessors()
     convert_controlnet()

src/refiners/foundationals/latent_diffusion/stable_diffusion_1/model.py

@@ -89,10 +89,18 @@ def compute_self_attention_guidance(
             classifier_free_guidance=True,
         )
 
-        negative_embedding, _ = clip_text_embedding.chunk(2)
         timestep = self.scheduler.timesteps[step].unsqueeze(dim=0)
+        negative_embedding, _ = clip_text_embedding.chunk(2)
         self.set_unet_context(timestep=timestep, clip_text_embedding=negative_embedding, **kwargs)
-        degraded_noise = self.unet(degraded_latents)
+        if "ip_adapter" in self.unet.provider.contexts:
+            # this implementation is a bit hacky, it should be refactored in the future
+            ip_adapter_context = self.unet.use_context("ip_adapter")
+            image_embedding_copy = ip_adapter_context["clip_image_embedding"].clone()
+            ip_adapter_context["clip_image_embedding"], _ = ip_adapter_context["clip_image_embedding"].chunk(2)
+            degraded_noise = self.unet(degraded_latents)
+            ip_adapter_context["clip_image_embedding"] = image_embedding_copy
+        else:
+            degraded_noise = self.unet(degraded_latents)
 
         return sag.scale * (noise - degraded_noise)
 
@@ -160,14 +168,23 @@ def compute_self_attention_guidance(
             step=step,
             classifier_free_guidance=True,
         )
-
-        negative_embedding, _ = clip_text_embedding.chunk(2)
-        timestep = self.scheduler.timesteps[step].unsqueeze(dim=0)
-        self.set_unet_context(timestep=timestep, clip_text_embedding=negative_embedding, **kwargs)
         x = torch.cat(
             tensors=(degraded_latents, self.mask_latents, self.target_image_latents),
             dim=1,
         )
-        degraded_noise = self.unet(x)
+
+        timestep = self.scheduler.timesteps[step].unsqueeze(dim=0)
+        negative_embedding, _ = clip_text_embedding.chunk(2)
+        self.set_unet_context(timestep=timestep, clip_text_embedding=negative_embedding, **kwargs)
+
+        if "ip_adapter" in self.unet.provider.contexts:
+            # this implementation is a bit hacky, it should be refactored in the future
+            ip_adapter_context = self.unet.use_context("ip_adapter")
+            image_embedding_copy = ip_adapter_context["clip_image_embedding"].clone()
+            ip_adapter_context["clip_image_embedding"], _ = ip_adapter_context["clip_image_embedding"].chunk(2)
+            degraded_noise = self.unet(x)
+            ip_adapter_context["clip_image_embedding"] = image_embedding_copy
+        else:
+            degraded_noise = self.unet(x)
 
         return sag.scale * (noise - degraded_noise)

src/refiners/foundationals/latent_diffusion/stable_diffusion_xl/model.py

@@ -138,17 +138,25 @@ def compute_self_attention_guidance(
             classifier_free_guidance=True,
         )
 
-        negative_embedding, _ = clip_text_embedding.chunk(2)
+        negative_text_embedding, _ = clip_text_embedding.chunk(2)
         negative_pooled_embedding, _ = pooled_text_embedding.chunk(2)
         timestep = self.scheduler.timesteps[step].unsqueeze(dim=0)
         time_ids, _ = time_ids.chunk(2)
+
         self.set_unet_context(
             timestep=timestep,
-            clip_text_embedding=negative_embedding,
+            clip_text_embedding=negative_text_embedding,
             pooled_text_embedding=negative_pooled_embedding,
             time_ids=time_ids,
-            **kwargs,
         )
-        degraded_noise = self.unet(degraded_latents)
+        if "ip_adapter" in self.unet.provider.contexts:
+            # this implementation is a bit hacky, it should be refactored in the future
+            ip_adapter_context = self.unet.use_context("ip_adapter")
+            image_embedding_copy = ip_adapter_context["clip_image_embedding"].clone()
+            ip_adapter_context["clip_image_embedding"], _ = ip_adapter_context["clip_image_embedding"].chunk(2)
+            degraded_noise = self.unet(degraded_latents)
+            ip_adapter_context["clip_image_embedding"] = image_embedding_copy
+        else:
+            degraded_noise = self.unet(degraded_latents)
 
         return sag.scale * (noise - degraded_noise)

tests/e2e/test_diffusion.py

@@ -242,6 +242,20 @@ def expected_freeu(ref_path: Path) -> Image.Image:
     return Image.open(fp=ref_path / "expected_freeu.png").convert(mode="RGB")
 
 
+@pytest.fixture
+def hello_world_assets(ref_path: Path) -> tuple[Image.Image, Image.Image, Image.Image, Image.Image]:
+    assets = Path(__file__).parent.parent.parent / "assets"
+    dropy = assets / "dropy_logo.png"
+    image_prompt = assets / "dragon_quest_slime.jpg"
+    condition_image = assets / "dropy_canny.png"
+    return (
+        Image.open(fp=dropy).convert(mode="RGB"),
+        Image.open(fp=image_prompt).convert(mode="RGB"),
+        Image.open(fp=condition_image).convert(mode="RGB"),
+        Image.open(fp=ref_path / "expected_dropy_slime_9752.png").convert(mode="RGB"),
+    )
+
+
 @pytest.fixture
 def text_embedding_textual_inversion(test_textual_inversion_path: Path) -> torch.Tensor:
     return torch.load(test_textual_inversion_path / "gta5-artwork" / "learned_embeds.bin")["<gta5-artwork>"]  # type: ignore
@@ -488,6 +502,15 @@ def sdxl_lda_weights(test_weights_path: Path) -> Path:
     return sdxl_lda_weights
 
 
+@pytest.fixture
+def sdxl_lda_fp16_fix_weights(test_weights_path: Path) -> Path:
+    sdxl_lda_weights = test_weights_path / "sdxl-lda-fp16-fix.safetensors"
+    if not sdxl_lda_weights.is_file():
+        warn(message=f"could not find weights at {sdxl_lda_weights}, skipping")
+        pytest.skip(allow_module_level=True)
+    return sdxl_lda_weights
+
+
 @pytest.fixture
 def sdxl_unet_weights(test_weights_path: Path) -> Path:
     sdxl_unet_weights = test_weights_path / "sdxl-unet.safetensors"
@@ -524,6 +547,24 @@ def sdxl_ddim(
     return sdxl
 
 
+@pytest.fixture
+def sdxl_ddim_lda_fp16_fix(
+    sdxl_text_encoder_weights: Path, sdxl_lda_fp16_fix_weights: Path, sdxl_unet_weights: Path, test_device: torch.device
+) -> StableDiffusion_XL:
+    if test_device.type == "cpu":
+        warn(message="not running on CPU, skipping")
+        pytest.skip()
+
+    scheduler = DDIM(num_inference_steps=30)
+    sdxl = StableDiffusion_XL(scheduler=scheduler, device=test_device)
+
+    sdxl.clip_text_encoder.load_from_safetensors(tensors_path=sdxl_text_encoder_weights)
+    sdxl.lda.load_from_safetensors(tensors_path=sdxl_lda_fp16_fix_weights)
+    sdxl.unet.load_from_safetensors(tensors_path=sdxl_unet_weights)
+
+    return sdxl
+
+
 @no_grad()
 def test_diffusion_std_random_init(
     sd15_std: StableDiffusion_1, expected_image_std_random_init: Image.Image, test_device: torch.device
@@ -1702,3 +1743,62 @@ def test_freeu(
     predicted_image = sd15.lda.decode_latents(x)
 
     ensure_similar_images(predicted_image, expected_freeu)
+
+
+@no_grad()
+def test_hello_world(
+    sdxl_ddim_lda_fp16_fix: StableDiffusion_XL,
+    t2i_adapter_xl_data_canny: tuple[str, Image.Image, Image.Image, Path],
+    sdxl_ip_adapter_weights: Path,
+    image_encoder_weights: Path,
+    hello_world_assets: tuple[Image.Image, Image.Image, Image.Image, Image.Image],
+) -> None:
+    sdxl = sdxl_ddim_lda_fp16_fix.to(dtype=torch.float16)
+    sdxl.dtype = torch.float16  # FIXME: should not be necessary
+
+    name, _, _, weights_path = t2i_adapter_xl_data_canny
+    init_image, image_prompt, condition_image, expected_image = hello_world_assets
+
+    if not weights_path.is_file():
+        warn(f"could not find weights at {weights_path}, skipping")
+        pytest.skip(allow_module_level=True)
+
+    ip_adapter = SDXLIPAdapter(target=sdxl.unet, weights=load_from_safetensors(sdxl_ip_adapter_weights))
+    ip_adapter.clip_image_encoder.load_from_safetensors(image_encoder_weights)
+    ip_adapter.inject()
+
+    image_embedding = ip_adapter.compute_clip_image_embedding(ip_adapter.preprocess_image(image_prompt))
+    ip_adapter.set_clip_image_embedding(image_embedding)
+
+    # Note: default text prompts for IP-Adapter
+    clip_text_embedding, pooled_text_embedding = sdxl.compute_clip_text_embedding(
+        text="best quality, high quality", negative_text="monochrome, lowres, bad anatomy, worst quality, low quality"
+    )
+    time_ids = sdxl.default_time_ids
+
+    t2i_adapter = SDXLT2IAdapter(target=sdxl.unet, name=name, weights=load_from_safetensors(weights_path)).inject()
+
+    condition = image_to_tensor(condition_image.convert("RGB"), device=sdxl.device, dtype=sdxl.dtype)
+    t2i_adapter.set_condition_features(features=t2i_adapter.compute_condition_features(condition))
+
+    first_step = 1
+    ip_adapter.set_scale(0.85)
+    t2i_adapter.set_scale(0.8)
+    sdxl.set_num_inference_steps(50)
+    sdxl.set_self_attention_guidance(enable=True, scale=0.75)
+
+    manual_seed(9752)
+    x = sdxl.init_latents(size=(1024, 1024), init_image=init_image, first_step=first_step).to(
+        device=sdxl.device, dtype=sdxl.dtype
+    )
+    for step in sdxl.steps[first_step:]:
+        x = sdxl(
+            x,
+            step=step,
+            clip_text_embedding=clip_text_embedding,
+            pooled_text_embedding=pooled_text_embedding,
+            time_ids=time_ids,
+        )
+    predicted_image = sdxl.lda.decode_latents(x)
+
+    ensure_similar_images(predicted_image, expected_image)

tests/e2e/test_diffusion_ref/README.md

@@ -47,6 +47,7 @@ Special cases:
     - `expected_cutecat_sdxl_ddim_random_init_sag.png`
     - `expected_restart.png`
     - `expected_freeu.png`
+    - `expected_dropy_slime_9752.png` (located in `/assets`)
 
 ## Other images