[tests] refactor l-series pipeline tests

src/diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion.py

@@ -113,7 +113,8 @@ def __call__(
                 generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                 tensor is generated by sampling using the supplied random `generator`.
             output_type (`str`, *optional*, defaults to `"pil"`):
-                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+                The output format of the generated image. Choose between `"pil"` (`PIL.Image`), `"np"` (`np.array`) or
+                `"pt"` (`torch.Tensor`).
             return_dict (`bool`, *optional*, defaults to `True`):
                 Whether or not to return a [`ImagePipelineOutput`] instead of a plain tuple.
 
@@ -220,9 +221,12 @@ def __call__(
         image = self.vqvae.decode(latents).sample
 
         image = (image / 2 + 0.5).clamp(0, 1)
-        image = image.cpu().permute(0, 2, 3, 1).numpy()
-        if output_type == "pil":
-            image = self.numpy_to_pil(image)
+
+        if output_type != "pt":
+            image = image.cpu().permute(0, 2, 3, 1).numpy()
+
+            if output_type == "pil":
+                image = self.numpy_to_pil(image)
 
         if not return_dict:
             return (image,)

src/diffusers/pipelines/latent_diffusion/pipeline_latent_diffusion_superresolution.py

@@ -97,7 +97,8 @@ def __call__(
                 A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make
                 generation deterministic.
             output_type (`str`, *optional*, defaults to `"pil"`):
-                The output format of the generated image. Choose between `PIL.Image` or `np.array`.
+                The output format of the generated image. Choose between `"pil"` (`PIL.Image`), `"np"` (`np.array`) or
+                `"pt"` (`torch.Tensor`).
             return_dict (`bool`, *optional*, defaults to `True`):
                 Whether or not to return a [`ImagePipelineOutput`] instead of a plain tuple.
 
@@ -185,10 +186,12 @@ def __call__(
         image = self.vqvae.decode(latents).sample
         image = torch.clamp(image, -1.0, 1.0)
         image = image / 2 + 0.5
-        image = image.cpu().permute(0, 2, 3, 1).numpy()
 
-        if output_type == "pil":
-            image = self.numpy_to_pil(image)
+        if output_type != "pt":
+            image = image.cpu().permute(0, 2, 3, 1).numpy()
+
+            if output_type == "pil":
+                image = self.numpy_to_pil(image)
 
         if not return_dict:
             return (image,)

tests/pipelines/latent_consistency_models/test_latent_consistency_models.py

@@ -1,8 +1,22 @@
+# coding=utf-8
+# Copyright 2026 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
 import gc
-import inspect
-import unittest
 
 import numpy as np
+import pytest
 import torch
 from transformers import CLIPTextConfig, CLIPTextModel, CLIPTokenizer
 
@@ -14,27 +28,43 @@
 )
 
 from ...testing_utils import (
+    assert_tensors_close,
     backend_empty_cache,
     enable_full_determinism,
     require_torch_accelerator,
     slow,
     torch_device,
 )
-from ..pipeline_params import TEXT_TO_IMAGE_BATCH_PARAMS, TEXT_TO_IMAGE_IMAGE_PARAMS, TEXT_TO_IMAGE_PARAMS
-from ..test_pipelines_common import IPAdapterTesterMixin, PipelineLatentTesterMixin, PipelineTesterMixin
+from ..testing_utils import (
+    BasePipelineTesterConfig,
+    IPAdapterTesterMixin,
+    LoraMemoryTesterMixin,
+    LoraTesterMixin,
+    MemoryTesterMixin,
+    PipelineTesterMixin,
+    UNetLoraTesterMixin,
+)
 
 
 enable_full_determinism()
 
 
-class LatentConsistencyModelPipelineFastTests(
-    IPAdapterTesterMixin, PipelineLatentTesterMixin, PipelineTesterMixin, unittest.TestCase
-):
+class LatentConsistencyModelPipelineTesterConfig(BasePipelineTesterConfig):
     pipeline_class = LatentConsistencyModelPipeline
-    params = TEXT_TO_IMAGE_PARAMS - {"negative_prompt", "negative_prompt_embeds"}
-    batch_params = TEXT_TO_IMAGE_BATCH_PARAMS - {"negative_prompt"}
-    image_params = TEXT_TO_IMAGE_IMAGE_PARAMS
-    image_latents_params = TEXT_TO_IMAGE_IMAGE_PARAMS
+    # The canonical text-to-image sets minus `negative_prompt` / `negative_prompt_embeds`: LCM is
+    # guidance-distilled and `__call__` takes no negative prompt.
+    required_input_params_in_call_signature = frozenset(
+        [
+            "prompt",
+            "height",
+            "width",
+            "guidance_scale",
+            "prompt_embeds",
+            "cross_attention_kwargs",
+        ]
+    )
+    batch_input_params = frozenset(["prompt"])
+    output_shape = (3, 64, 64)
 
     def get_dummy_components(self):
         torch.manual_seed(0)
@@ -82,7 +112,7 @@ def get_dummy_components(self):
         text_encoder = CLIPTextModel(text_encoder_config)
         tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
 
-        components = {
+        return {
             "unet": unet,
             "scheduler": scheduler,
             "vae": vae,
@@ -93,155 +123,144 @@ def get_dummy_components(self):
             "image_encoder": None,
             "requires_safety_checker": False,
         }
-        return components
-
-    def get_dummy_inputs(self, device, seed=0):
-        if str(device).startswith("mps"):
-            generator = torch.manual_seed(seed)
-        else:
-            generator = torch.Generator(device=device).manual_seed(seed)
-        inputs = {
+
+    def get_dummy_inputs(self):
+        return {
             "prompt": "A painting of a squirrel eating a burger",
-            "generator": generator,
+            "generator": self.get_generator(0),
             "num_inference_steps": 2,
             "guidance_scale": 6.0,
-            "output_type": "np",
+            # Request torch outputs so tests compare torch tensors directly (see `BasePipelineTesterConfig`).
+            # Note `"pt"` images are `(batch, channels, height, width)`, unlike `"np"` (`(batch, h, w, c)`).
+            "output_type": "pt",
         }
-        return inputs
 
-    def test_ip_adapter(self):
-        expected_pipe_slice = None
-        if torch_device == "cpu":
-            expected_pipe_slice = np.array([0.1405, 0.5002, 0.5213, 0.1223, 0.3856, 0.4165, 0.5382, 0.3622, 0.3693])
-        return super().test_ip_adapter(expected_pipe_slice=expected_pipe_slice)
 
+class TestLatentConsistencyModelPipeline(LatentConsistencyModelPipelineTesterConfig, PipelineTesterMixin):
     def test_lcm_onestep(self):
-        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
-
-        components = self.get_dummy_components()
-        pipe = LatentConsistencyModelPipeline(**components)
-        pipe = pipe.to(torch_device)
-        pipe.set_progress_bar_config(disable=None)
+        # Run on CPU: the expected slice below is CPU-specific.
+        pipe = self.get_pipeline()
 
-        inputs = self.get_dummy_inputs(device)
+        inputs = self.get_dummy_inputs()
         inputs["num_inference_steps"] = 1
-        output = pipe(**inputs)
-        image = output.images
-        assert image.shape == (1, 64, 64, 3)
+        image = pipe(**inputs).images
+        assert image.shape == (1, *self.output_shape)
 
-        image_slice = image[0, -3:, -3:, -1]
-        expected_slice = np.array([0.1444, 0.5229, 0.5344, 0.1384, 0.3999, 0.4320, 0.5345, 0.3559, 0.3685])
-        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
+        # fmt: off
+        expected_slice = torch.tensor([0.1444, 0.5229, 0.5344, 0.1384, 0.3999, 0.4320, 0.5345, 0.3559, 0.3685])
+        # fmt: on
+        assert_tensors_close(image[0, -1, -3:, -3:].flatten(), expected_slice, atol=1e-3)
 
     def test_lcm_multistep(self):
-        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
-
-        components = self.get_dummy_components()
-        pipe = LatentConsistencyModelPipeline(**components)
-        pipe = pipe.to(torch_device)
-        pipe.set_progress_bar_config(disable=None)
+        # Run on CPU: the expected slice below is CPU-specific.
+        pipe = self.get_pipeline()
 
-        inputs = self.get_dummy_inputs(device)
-        output = pipe(**inputs)
-        image = output.images
-        assert image.shape == (1, 64, 64, 3)
+        image = pipe(**self.get_dummy_inputs()).images
+        assert image.shape == (1, *self.output_shape)
 
-        image_slice = image[0, -3:, -3:, -1]
-        expected_slice = np.array([0.1405, 0.5002, 0.5213, 0.1223, 0.3856, 0.4165, 0.5382, 0.3622, 0.3693])
-        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
+        # fmt: off
+        expected_slice = torch.tensor([0.1405, 0.5002, 0.5213, 0.1223, 0.3856, 0.4165, 0.5382, 0.3622, 0.3693])
+        # fmt: on
+        assert_tensors_close(image[0, -1, -3:, -3:].flatten(), expected_slice, atol=1e-3)
 
     def test_lcm_custom_timesteps(self):
-        device = "cpu"  # ensure determinism for the device-dependent torch.Generator
+        # Run on CPU: the expected slice below is CPU-specific.
+        pipe = self.get_pipeline()
 
-        components = self.get_dummy_components()
-        pipe = LatentConsistencyModelPipeline(**components)
-        pipe = pipe.to(torch_device)
-        pipe.set_progress_bar_config(disable=None)
-
-        inputs = self.get_dummy_inputs(device)
+        inputs = self.get_dummy_inputs()
         del inputs["num_inference_steps"]
         inputs["timesteps"] = [999, 499]
-        output = pipe(**inputs)
-        image = output.images
-        assert image.shape == (1, 64, 64, 3)
+        image = pipe(**inputs).images
+        assert image.shape == (1, *self.output_shape)
 
-        image_slice = image[0, -3:, -3:, -1]
-        expected_slice = np.array([0.1405, 0.5002, 0.5213, 0.1223, 0.3856, 0.4165, 0.5382, 0.3622, 0.3693])
-        assert np.abs(image_slice.flatten() - expected_slice).max() < 1e-3
+        # Custom timesteps matching the default 2-step schedule reproduce `test_lcm_multistep`'s output.
+        # fmt: off
+        expected_slice = torch.tensor([0.1405, 0.5002, 0.5213, 0.1223, 0.3856, 0.4165, 0.5382, 0.3622, 0.3693])
+        # fmt: on
+        assert_tensors_close(image[0, -1, -3:, -3:].flatten(), expected_slice, atol=1e-3)
 
-    def test_inference_batch_single_identical(self):
-        super().test_inference_batch_single_identical(expected_max_diff=5e-4)
+    def test_inference_batch_single_identical(self, batch_size=3, expected_max_diff=5e-4):
+        super().test_inference_batch_single_identical(batch_size=batch_size, expected_max_diff=expected_max_diff)
 
-    # skip because lcm pipeline apply cfg differently
+    @pytest.mark.skip("LCM applies classifier-free guidance differently, so the shared CFG callback test cannot run.")
     def test_callback_cfg(self):
         pass
 
-    # override default test because the final latent variable is "denoised" instead of "latents"
     def test_callback_inputs(self):
-        sig = inspect.signature(self.pipeline_class.__call__)
+        # Overridden because the final latent variable is `denoised` rather than `latents`.
+        pipe = self.get_pipeline().to(torch_device)
 
-        if not ("callback_on_step_end_tensor_inputs" in sig.parameters and "callback_on_step_end" in sig.parameters):
-            return
-
-        components = self.get_dummy_components()
-        pipe = self.pipeline_class(**components)
-        pipe = pipe.to(torch_device)
-        pipe.set_progress_bar_config(disable=None)
-
-        self.assertTrue(
-            hasattr(pipe, "_callback_tensor_inputs"),
-            f" {self.pipeline_class} should have `_callback_tensor_inputs` that defines a list of tensor variables its callback function can use as inputs",
+        assert hasattr(pipe, "_callback_tensor_inputs"), (
+            f"{self.pipeline_class} should have `_callback_tensor_inputs` that defines a list of tensor variables "
+            "its callback function can use as inputs"
         )
 
         def callback_inputs_test(pipe, i, t, callback_kwargs):
-            missing_callback_inputs = set()
-            for v in pipe._callback_tensor_inputs:
-                if v not in callback_kwargs:
-                    missing_callback_inputs.add(v)
-            self.assertTrue(
-                len(missing_callback_inputs) == 0, f"Missing callback tensor inputs: {missing_callback_inputs}"
-            )
-            last_i = pipe.num_timesteps - 1
-            if i == last_i:
+            missing_callback_inputs = {v for v in pipe._callback_tensor_inputs if v not in callback_kwargs}
+            assert len(missing_callback_inputs) == 0, f"Missing callback tensor inputs: {missing_callback_inputs}"
+            if i == pipe.num_timesteps - 1:
                 callback_kwargs["denoised"] = torch.zeros_like(callback_kwargs["denoised"])
             return callback_kwargs
 
-        inputs = self.get_dummy_inputs(torch_device)
+        inputs = self.get_dummy_inputs()
         inputs["callback_on_step_end"] = callback_inputs_test
         inputs["callback_on_step_end_tensor_inputs"] = pipe._callback_tensor_inputs
         inputs["output_type"] = "latent"
 
         output = pipe(**inputs)[0]
         assert output.abs().sum() == 0
 
-    def test_encode_prompt_works_in_isolation(self):
+    def test_encode_prompt_works_in_isolation(self, extra_required_param_value_dict=None, atol=1e-4, rtol=1e-4):
+        # `encode_prompt` requires `device` and `do_classifier_free_guidance`, neither of which `__call__`
+        # exposes with a default for the shared test to pick up.
         extra_required_param_value_dict = {
             "device": torch.device(torch_device).type,
-            "do_classifier_free_guidance": self.get_dummy_inputs(device=torch_device).get("guidance_scale", 1.0) > 1.0,
+            "do_classifier_free_guidance": self.get_dummy_inputs().get("guidance_scale", 1.0) > 1.0,
         }
-        return super().test_encode_prompt_works_in_isolation(extra_required_param_value_dict)
+        super().test_encode_prompt_works_in_isolation(extra_required_param_value_dict, atol=atol, rtol=rtol)
+
+
+class TestLatentConsistencyModelPipelineMemory(LatentConsistencyModelPipelineTesterConfig, MemoryTesterMixin):
+    """Memory optimization tests (CPU offload, group offload, layerwise casting) for the LCM pipeline."""
+
+
+class TestLatentConsistencyModelPipelineIPAdapter(LatentConsistencyModelPipelineTesterConfig, IPAdapterTesterMixin):
+    """IP-Adapter tests for the LCM pipeline."""
+
+
+class TestLatentConsistencyModelPipelineLoRA(
+    LatentConsistencyModelPipelineTesterConfig, LoraTesterMixin, UNetLoraTesterMixin
+):
+    """LoRA tests for the LCM pipeline."""
+
+
+class TestLatentConsistencyModelPipelineLoRAMemory(LatentConsistencyModelPipelineTesterConfig, LoraMemoryTesterMixin):
+    """LoRA x memory-optimization tests (group offload, CPU offload) for the LCM pipeline."""
 
 
 @slow
 @require_torch_accelerator
-class LatentConsistencyModelPipelineSlowTests(unittest.TestCase):
-    def setUp(self):
+class TestLatentConsistencyModelPipelineIntegration:
+    @pytest.fixture(autouse=True)
+    def cleanup(self):
+        gc.collect()
+        backend_empty_cache(torch_device)
+        yield
         gc.collect()
         backend_empty_cache(torch_device)
 
     def get_inputs(self, device, generator_device="cpu", dtype=torch.float32, seed=0):
         generator = torch.Generator(device=generator_device).manual_seed(seed)
         latents = np.random.RandomState(seed).standard_normal((1, 4, 64, 64))
         latents = torch.from_numpy(latents).to(device=device, dtype=dtype)
-        inputs = {
+        return {
             "prompt": "a photograph of an astronaut riding a horse",
             "latents": latents,
             "generator": generator,
             "num_inference_steps": 3,
             "guidance_scale": 7.5,
             "output_type": "np",
         }
-        return inputs
 
     def test_lcm_onestep(self):
         pipe = LatentConsistencyModelPipeline.from_pretrained("SimianLuo/LCM_Dreamshaper_v7", safety_checker=None)
@@ -264,8 +283,7 @@ def test_lcm_multistep(self):
         pipe = pipe.to(torch_device)
         pipe.set_progress_bar_config(disable=None)
 
-        inputs = self.get_inputs(torch_device)
-        image = pipe(**inputs).images
+        image = pipe(**self.get_inputs(torch_device)).images
         assert image.shape == (1, 512, 512, 3)
 
         image_slice = image[0, -3:, -3:, -1].flatten()