Bria 3 2 pipeline (#12010)

* Add Bria model and pipeline to diffusers

- Introduced `BriaTransformer2DModel` and `BriaPipeline` for enhanced image generation capabilities.
- Updated import structures across various modules to include the new Bria components.
- Added utility functions and output classes specific to the Bria pipeline.
- Implemented tests for the Bria pipeline to ensure functionality and output integrity.

* with working tests

* style and quality pass

* adding docs

* add to overview

* fixes from "make fix-copies"

* Refactor transformer_bria.py and pipeline_bria.py: Introduce new EmbedND class for rotary position embedding, and enhance Timestep and TimestepProjEmbeddings classes. Add utility functions for handling negative prompts and generating original sigmas in pipeline_bria.py.

* remove redundent and duplicates tests and fix bf16
slow test

* style fixes

* small doc update

* Enhance Bria 3.2 documentation and implementation

- Updated the GitHub repository link for Bria 3.2.
- Added usage instructions for the gated model access.
- Introduced the BriaTransformerBlock and BriaAttention classes to the model architecture.
- Refactored existing classes to integrate Bria-specific components, including BriaEmbedND and BriaPipeline.
- Updated the pipeline output class to reflect Bria-specific functionality.
- Adjusted test cases to align with the new Bria model structure.

* Refactor Bria model components and update documentation

- Removed outdated inference example from Bria 3.2 documentation.
- Introduced the BriaTransformerBlock class to enhance model architecture.
- Updated attention handling to use `attention_kwargs` instead of `joint_attention_kwargs`.
- Improved import structure in the Bria pipeline to handle optional dependencies.
- Adjusted test cases to reflect changes in model dtype assertions.

* Update Bria model reference in documentation to reflect new file naming convention

* Update docs/source/en/_toctree.yml

* Refactor BriaPipeline to inherit from DiffusionPipeline instead of FluxPipeline, updating imports accordingly.

* move the __call__ func to the end of file

* Update BriaPipeline example to use bfloat16 for precision sensitivity for better result

* make style && make quality &&  make fix-copiessource

---------
Co-authored-by: Linoy Tsaban <57615435+linoytsaban@users.noreply.github.com>
Co-authored-by: Aryan <contact.aryanvs@gmail.com>

docs/source/en/_toctree.yml

@@ -340,6 +340,8 @@
         title: AllegroTransformer3DModel
       - local: api/models/aura_flow_transformer2d
         title: AuraFlowTransformer2DModel
+      - local: api/models/bria_transformer
+        title: BriaTransformer2DModel
       - local: api/models/chroma_transformer
         title: ChromaTransformer2DModel
       - local: api/models/cogvideox_transformer3d
@@ -468,6 +470,8 @@
       title: AutoPipeline
     - local: api/pipelines/blip_diffusion
       title: BLIP-Diffusion
+    - local: api/pipelines/bria_3_2
+      title: Bria 3.2
     - local: api/pipelines/chroma
       title: Chroma
     - local: api/pipelines/cogvideox

docs/source/en/api/models/bria_transformer.md

+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+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.
+-->
+
+# BriaTransformer2DModel
+
+A modified flux Transformer model from [Bria](https://huggingface.co/briaai/BRIA-3.2)
+
+## BriaTransformer2DModel
+
+[[autodoc]] BriaTransformer2DModel

docs/source/en/api/pipelines/bria_3_2.md

+<!--Copyright 2025 The HuggingFace Team. All rights reserved.
+
+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.
+-->
+
+# Bria 3.2
+
+Bria 3.2 is the next-generation commercial-ready text-to-image model. With just 4 billion parameters, it provides exceptional aesthetics and text rendering, evaluated to provide on par results to leading open-source models, and outperforming other licensed models.
+In addition to being built entirely on licensed data, 3.2 provides several advantages for enterprise and commercial use:
+
+- Efficient Compute - the model is X3 smaller than the equivalent models in the market (4B parameters vs 12B parameters other open source models)
+- Architecture Consistency: Same architecture as 3.1—ideal for users looking to upgrade without disruption.
+- Fine-tuning Speedup: 2x faster fine-tuning on L40S and A100.
+
+Original model checkpoints for Bria 3.2 can be found [here](https://huggingface.co/briaai/BRIA-3.2).
+Github repo for Bria 3.2 can be found [here](https://github.com/Bria-AI/BRIA-3.2).
+
+If you want to learn more about the Bria platform, and get free traril access, please visit [bria.ai](https://bria.ai).
+
+
+## Usage
+
+_As the model is gated, before using it with diffusers you first need to go to the [Bria 3.2 Hugging Face page](https://huggingface.co/briaai/BRIA-3.2), fill in the form and accept the gate. Once you are in, you need to login so that your system knows you’ve accepted the gate._
+
+Use the command below to log in:
+
+```bash
+hf auth login
+```
+
+
+## BriaPipeline
+
+[[autodoc]] BriaPipeline
+	- all
+	- __call__
+

docs/source/en/api/pipelines/overview.md

@@ -37,6 +37,7 @@ The table below lists all the pipelines currently available in 🤗 Diffusers an
 | [AudioLDM2](audioldm2) | text2audio |
 | [AuraFlow](auraflow) | text2image |
 | [BLIP Diffusion](blip_diffusion) | text2image |
+| [Bria 3.2](bria_3_2) | text2image |
 | [CogVideoX](cogvideox) | text2video |
 | [Consistency Models](consistency_models) | unconditional image generation |
 | [ControlNet](controlnet) | text2image, image2image, inpainting |

src/diffusers/__init__.py

@@ -181,6 +181,7 @@ else:
             "AutoencoderOobleck",
             "AutoencoderTiny",
             "AutoModel",
+            "BriaTransformer2DModel",
             "CacheMixin",
             "ChromaTransformer2DModel",
             "CogVideoXTransformer3DModel",
@@ -397,6 +398,7 @@ else:
             "AuraFlowPipeline",
             "BlipDiffusionControlNetPipeline",
             "BlipDiffusionPipeline",
+            "BriaPipeline",
             "ChromaImg2ImgPipeline",
             "ChromaPipeline",
             "CLIPImageProjection",
@@ -846,6 +848,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AutoencoderOobleck,
             AutoencoderTiny,
             AutoModel,
+            BriaTransformer2DModel,
             CacheMixin,
             ChromaTransformer2DModel,
             CogVideoXTransformer3DModel,
@@ -1032,6 +1035,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             AudioLDM2UNet2DConditionModel,
             AudioLDMPipeline,
             AuraFlowPipeline,
+            BriaPipeline,
             ChromaImg2ImgPipeline,
             ChromaPipeline,
             CLIPImageProjection,

src/diffusers/hooks/_helpers.py

@@ -144,6 +144,7 @@ def _register_attention_processors_metadata():
 def _register_transformer_blocks_metadata():
     from ..models.attention import BasicTransformerBlock
     from ..models.transformers.cogvideox_transformer_3d import CogVideoXBlock
+    from ..models.transformers.transformer_bria import BriaTransformerBlock
     from ..models.transformers.transformer_cogview4 import CogView4TransformerBlock
     from ..models.transformers.transformer_flux import FluxSingleTransformerBlock, FluxTransformerBlock
     from ..models.transformers.transformer_hunyuan_video import (
@@ -165,6 +166,13 @@ def _register_transformer_blocks_metadata():
             return_encoder_hidden_states_index=None,
         ),
     )
+    TransformerBlockRegistry.register(
+        model_class=BriaTransformerBlock,
+        metadata=TransformerBlockMetadata(
+            return_hidden_states_index=0,
+            return_encoder_hidden_states_index=None,
+        ),
+    )
 
     # CogVideoX
     TransformerBlockRegistry.register(

src/diffusers/models/__init__.py

@@ -76,6 +76,7 @@ if is_torch_available():
     _import_structure["transformers.t5_film_transformer"] = ["T5FilmDecoder"]
     _import_structure["transformers.transformer_2d"] = ["Transformer2DModel"]
     _import_structure["transformers.transformer_allegro"] = ["AllegroTransformer3DModel"]
+    _import_structure["transformers.transformer_bria"] = ["BriaTransformer2DModel"]
     _import_structure["transformers.transformer_chroma"] = ["ChromaTransformer2DModel"]
     _import_structure["transformers.transformer_cogview3plus"] = ["CogView3PlusTransformer2DModel"]
     _import_structure["transformers.transformer_cogview4"] = ["CogView4Transformer2DModel"]
@@ -158,6 +159,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         from .transformers import (
             AllegroTransformer3DModel,
             AuraFlowTransformer2DModel,
+            BriaTransformer2DModel,
             ChromaTransformer2DModel,
             CogVideoXTransformer3DModel,
             CogView3PlusTransformer2DModel,

src/diffusers/models/transformers/__init__.py

@@ -17,6 +17,7 @@ if is_torch_available():
     from .t5_film_transformer import T5FilmDecoder
     from .transformer_2d import Transformer2DModel
     from .transformer_allegro import AllegroTransformer3DModel
+    from .transformer_bria import BriaTransformer2DModel
     from .transformer_chroma import ChromaTransformer2DModel
     from .transformer_cogview3plus import CogView3PlusTransformer2DModel
     from .transformer_cogview4 import CogView4Transformer2DModel

src/diffusers/pipelines/__init__.py

@@ -127,6 +127,7 @@ else:
         "AnimateDiffVideoToVideoPipeline",
         "AnimateDiffVideoToVideoControlNetPipeline",
     ]
+    _import_structure["bria"] = ["BriaPipeline"]
     _import_structure["flux"] = [
         "FluxControlPipeline",
         "FluxControlInpaintPipeline",
@@ -552,6 +553,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
         )
         from .aura_flow import AuraFlowPipeline
         from .blip_diffusion import BlipDiffusionPipeline
+        from .bria import BriaPipeline
         from .chroma import ChromaImg2ImgPipeline, ChromaPipeline
         from .cogvideo import (
             CogVideoXFunControlPipeline,

src/diffusers/pipelines/bria/__init__.py

+from typing import TYPE_CHECKING
+
+from ...utils import (
+    DIFFUSERS_SLOW_IMPORT,
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    get_objects_from_module,
+    is_torch_available,
+    is_transformers_available,
+)
+
+
+_dummy_objects = {}
+_import_structure = {}
+
+
+try:
+    if not (is_transformers_available() and is_torch_available()):
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    from ...utils import dummy_torch_and_transformers_objects  # noqa F403
+
+    _dummy_objects.update(get_objects_from_module(dummy_torch_and_transformers_objects))
+else:
+    _import_structure["pipeline_bria"] = ["BriaPipeline"]
+
+if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
+    try:
+        if not (is_transformers_available() and is_torch_available()):
+            raise OptionalDependencyNotAvailable()
+
+    except OptionalDependencyNotAvailable:
+        from ...utils.dummy_torch_and_transformers_objects import *
+    else:
+        from .pipeline_bria import BriaPipeline
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(
+        __name__,
+        globals()["__file__"],
+        _import_structure,
+        module_spec=__spec__,
+    )
+
+    for name, value in _dummy_objects.items():
+        setattr(sys.modules[__name__], name, value)

src/diffusers/pipelines/bria/pipeline_output.py

+from dataclasses import dataclass
+from typing import List, Union
+
+import numpy as np
+import PIL.Image
+
+from ...utils import BaseOutput
+
+
+@dataclass
+class BriaPipelineOutput(BaseOutput):
+    """
+    Output class for Bria pipelines.
+
+    Args:
+        images (`List[PIL.Image.Image]` or `np.ndarray`)
+            List of denoised PIL images of length `batch_size` or numpy array of shape `(batch_size, height, width,
+            num_channels)`. PIL images or numpy array present the denoised images of the diffusion pipeline.
+    """
+
+    images: Union[List[PIL.Image.Image], np.ndarray]

src/diffusers/utils/dummy_pt_objects.py

@@ -528,6 +528,21 @@ class AutoModel(metaclass=DummyObject):
         requires_backends(cls, ["torch"])
 
 
+class BriaTransformer2DModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+
 class CacheMixin(metaclass=DummyObject):
     _backends = ["torch"]
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -362,6 +362,21 @@ class AuraFlowPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class BriaPipeline(metaclass=DummyObject):
+    _backends = ["torch", "transformers"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch", "transformers"])
+
+    @classmethod
+    def from_config(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch", "transformers"])
+
+
 class ChromaImg2ImgPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/models/transformers/test_models_transformer_bria.py

+# coding=utf-8
+# Copyright 2025 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 unittest
+
+import torch
+
+from diffusers import BriaTransformer2DModel
+from diffusers.models.attention_processor import FluxIPAdapterJointAttnProcessor2_0
+from diffusers.models.embeddings import ImageProjection
+from diffusers.utils.testing_utils import enable_full_determinism, torch_device
+
+from ..test_modeling_common import LoraHotSwappingForModelTesterMixin, ModelTesterMixin, TorchCompileTesterMixin
+
+
+enable_full_determinism()
+
+
+def create_bria_ip_adapter_state_dict(model):
+    # "ip_adapter" (cross-attention weights)
+    ip_cross_attn_state_dict = {}
+    key_id = 0
+
+    for name in model.attn_processors.keys():
+        if name.startswith("single_transformer_blocks"):
+            continue
+
+        joint_attention_dim = model.config["joint_attention_dim"]
+        hidden_size = model.config["num_attention_heads"] * model.config["attention_head_dim"]
+        sd = FluxIPAdapterJointAttnProcessor2_0(
+            hidden_size=hidden_size, cross_attention_dim=joint_attention_dim, scale=1.0
+        ).state_dict()
+        ip_cross_attn_state_dict.update(
+            {
+                f"{key_id}.to_k_ip.weight": sd["to_k_ip.0.weight"],
+                f"{key_id}.to_v_ip.weight": sd["to_v_ip.0.weight"],
+                f"{key_id}.to_k_ip.bias": sd["to_k_ip.0.bias"],
+                f"{key_id}.to_v_ip.bias": sd["to_v_ip.0.bias"],
+            }
+        )
+
+        key_id += 1
+
+    # "image_proj" (ImageProjection layer weights)
+
+    image_projection = ImageProjection(
+        cross_attention_dim=model.config["joint_attention_dim"],
+        image_embed_dim=model.config["pooled_projection_dim"],
+        num_image_text_embeds=4,
+    )
+
+    ip_image_projection_state_dict = {}
+    sd = image_projection.state_dict()
+    ip_image_projection_state_dict.update(
+        {
+            "proj.weight": sd["image_embeds.weight"],
+            "proj.bias": sd["image_embeds.bias"],
+            "norm.weight": sd["norm.weight"],
+            "norm.bias": sd["norm.bias"],
+        }
+    )
+
+    del sd
+    ip_state_dict = {}
+    ip_state_dict.update({"image_proj": ip_image_projection_state_dict, "ip_adapter": ip_cross_attn_state_dict})
+    return ip_state_dict
+
+
+class BriaTransformerTests(ModelTesterMixin, unittest.TestCase):
+    model_class = BriaTransformer2DModel
+    main_input_name = "hidden_states"
+    # We override the items here because the transformer under consideration is small.
+    model_split_percents = [0.8, 0.7, 0.7]
+
+    # Skip setting testing with default: AttnProcessor
+    uses_custom_attn_processor = True
+
+    @property
+    def dummy_input(self):
+        batch_size = 1
+        num_latent_channels = 4
+        num_image_channels = 3
+        height = width = 4
+        sequence_length = 48
+        embedding_dim = 32
+
+        hidden_states = torch.randn((batch_size, height * width, num_latent_channels)).to(torch_device)
+        encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device)
+        text_ids = torch.randn((sequence_length, num_image_channels)).to(torch_device)
+        image_ids = torch.randn((height * width, num_image_channels)).to(torch_device)
+        timestep = torch.tensor([1.0]).to(torch_device).expand(batch_size)
+
+        return {
+            "hidden_states": hidden_states,
+            "encoder_hidden_states": encoder_hidden_states,
+            "img_ids": image_ids,
+            "txt_ids": text_ids,
+            "timestep": timestep,
+        }
+
+    @property
+    def input_shape(self):
+        return (16, 4)
+
+    @property
+    def output_shape(self):
+        return (16, 4)
+
+    def prepare_init_args_and_inputs_for_common(self):
+        init_dict = {
+            "patch_size": 1,
+            "in_channels": 4,
+            "num_layers": 1,
+            "num_single_layers": 1,
+            "attention_head_dim": 8,
+            "num_attention_heads": 2,
+            "joint_attention_dim": 32,
+            "pooled_projection_dim": None,
+            "axes_dims_rope": [0, 4, 4],
+        }
+
+        inputs_dict = self.dummy_input
+        return init_dict, inputs_dict
+
+    def test_deprecated_inputs_img_txt_ids_3d(self):
+        init_dict, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+        model = self.model_class(**init_dict)
+        model.to(torch_device)
+        model.eval()
+
+        with torch.no_grad():
+            output_1 = model(**inputs_dict).to_tuple()[0]
+
+        # update inputs_dict with txt_ids and img_ids as 3d tensors (deprecated)
+        text_ids_3d = inputs_dict["txt_ids"].unsqueeze(0)
+        image_ids_3d = inputs_dict["img_ids"].unsqueeze(0)
+
+        assert text_ids_3d.ndim == 3, "text_ids_3d should be a 3d tensor"
+        assert image_ids_3d.ndim == 3, "img_ids_3d should be a 3d tensor"
+
+        inputs_dict["txt_ids"] = text_ids_3d
+        inputs_dict["img_ids"] = image_ids_3d
+
+        with torch.no_grad():
+            output_2 = model(**inputs_dict).to_tuple()[0]
+
+        self.assertEqual(output_1.shape, output_2.shape)
+        self.assertTrue(
+            torch.allclose(output_1, output_2, atol=1e-5),
+            msg="output with deprecated inputs (img_ids and txt_ids as 3d torch tensors) are not equal as them as 2d inputs",
+        )
+
+    def test_gradient_checkpointing_is_applied(self):
+        expected_set = {"BriaTransformer2DModel"}
+        super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
+
+
+class BriaTransformerCompileTests(TorchCompileTesterMixin, unittest.TestCase):
+    model_class = BriaTransformer2DModel
+
+    def prepare_init_args_and_inputs_for_common(self):
+        return BriaTransformerTests().prepare_init_args_and_inputs_for_common()
+
+
+class BriaTransformerLoRAHotSwapTests(LoraHotSwappingForModelTesterMixin, unittest.TestCase):
+    model_class = BriaTransformer2DModel
+
+    def prepare_init_args_and_inputs_for_common(self):
+        return BriaTransformerTests().prepare_init_args_and_inputs_for_common()

tests/pipelines/bria/test_pipeline_bria.py

+# Copyright 2024 Bria AI and The HuggingFace Team. All rights reserved.
+#
+# 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 tempfile
+import unittest
+
+import numpy as np
+import torch
+from huggingface_hub import hf_hub_download
+from transformers import T5EncoderModel, T5TokenizerFast
+
+from diffusers import (
+    AutoencoderKL,
+    BriaTransformer2DModel,
+    FlowMatchEulerDiscreteScheduler,
+)
+from diffusers.pipelines.bria import BriaPipeline
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    numpy_cosine_similarity_distance,
+    require_accelerator,
+    require_torch_gpu,
+    slow,
+    torch_device,
+)
+
+# from ..test_pipelines_common import PipelineTesterMixin, check_qkv_fused_layers_exist
+from tests.pipelines.test_pipelines_common import PipelineTesterMixin, to_np
+
+
+enable_full_determinism()
+
+
+class BriaPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
+    pipeline_class = BriaPipeline
+    params = frozenset(["prompt", "height", "width", "guidance_scale", "prompt_embeds"])
+    batch_params = frozenset(["prompt"])
+    test_xformers_attention = False
+
+    # there is no xformers processor for Flux
+    test_xformers_attention = False
+    test_layerwise_casting = True
+    test_group_offloading = True
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        transformer = BriaTransformer2DModel(
+            patch_size=1,
+            in_channels=16,
+            num_layers=1,
+            num_single_layers=1,
+            attention_head_dim=8,
+            num_attention_heads=2,
+            joint_attention_dim=32,
+            pooled_projection_dim=None,
+            axes_dims_rope=[0, 4, 4],
+        )
+
+        torch.manual_seed(0)
+        vae = AutoencoderKL(
+            act_fn="silu",
+            block_out_channels=(32,),
+            in_channels=3,
+            out_channels=3,
+            down_block_types=["DownEncoderBlock2D"],
+            up_block_types=["UpDecoderBlock2D"],
+            latent_channels=4,
+            sample_size=32,
+            shift_factor=0,
+            scaling_factor=0.13025,
+            use_post_quant_conv=True,
+            use_quant_conv=True,
+            force_upcast=False,
+        )
+
+        scheduler = FlowMatchEulerDiscreteScheduler()
+
+        torch.manual_seed(0)
+        text_encoder = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+        tokenizer = T5TokenizerFast.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        components = {
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+            "transformer": transformer,
+            "vae": vae,
+            "image_encoder": None,
+            "feature_extractor": None,
+        }
+        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="cpu").manual_seed(seed)
+
+        inputs = {
+            "prompt": "A painting of a squirrel eating a burger",
+            "negative_prompt": "bad, ugly",
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 5.0,
+            "height": 16,
+            "width": 16,
+            "max_sequence_length": 48,
+            "output_type": "np",
+        }
+        return inputs
+
+    def test_encode_prompt_works_in_isolation(self):
+        pass
+
+    def test_bria_different_prompts(self):
+        pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device)
+        inputs = self.get_dummy_inputs(torch_device)
+        output_same_prompt = pipe(**inputs).images[0]
+        inputs = self.get_dummy_inputs(torch_device)
+        inputs["prompt"] = "a different prompt"
+        output_different_prompts = pipe(**inputs).images[0]
+        max_diff = np.abs(output_same_prompt - output_different_prompts).max()
+        assert max_diff > 1e-6
+
+    def test_image_output_shape(self):
+        pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device)
+        inputs = self.get_dummy_inputs(torch_device)
+
+        height_width_pairs = [(32, 32), (72, 57)]
+        for height, width in height_width_pairs:
+            expected_height = height - height % (pipe.vae_scale_factor * 2)
+            expected_width = width - width % (pipe.vae_scale_factor * 2)
+
+            inputs.update({"height": height, "width": width})
+            image = pipe(**inputs).images[0]
+            output_height, output_width, _ = image.shape
+            assert (output_height, output_width) == (expected_height, expected_width)
+
+    @unittest.skipIf(torch_device not in ["cuda", "xpu"], reason="float16 requires CUDA or XPU")
+    @require_accelerator
+    def test_save_load_float16(self, expected_max_diff=1e-2):
+        components = self.get_dummy_components()
+        for name, module in components.items():
+            if hasattr(module, "half"):
+                components[name] = module.to(torch_device).half()
+
+        pipe = self.pipeline_class(**components)
+        for component in pipe.components.values():
+            if hasattr(component, "set_default_attn_processor"):
+                component.set_default_attn_processor()
+        pipe.to(torch_device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(torch_device)
+        output = pipe(**inputs)[0]
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            pipe.save_pretrained(tmpdir)
+            pipe_loaded = self.pipeline_class.from_pretrained(tmpdir, torch_dtype=torch.float16)
+            for component in pipe_loaded.components.values():
+                if hasattr(component, "set_default_attn_processor"):
+                    component.set_default_attn_processor()
+            pipe_loaded.to(torch_device)
+            pipe_loaded.set_progress_bar_config(disable=None)
+
+        for name, component in pipe_loaded.components.items():
+            if name == "vae":
+                continue
+            if hasattr(component, "dtype"):
+                self.assertTrue(
+                    component.dtype == torch.float16,
+                    f"`{name}.dtype` switched from `float16` to {component.dtype} after loading.",
+                )
+
+        inputs = self.get_dummy_inputs(torch_device)
+        output_loaded = pipe_loaded(**inputs)[0]
+        max_diff = np.abs(to_np(output) - to_np(output_loaded)).max()
+        self.assertLess(
+            max_diff, expected_max_diff, "The output of the fp16 pipeline changed after saving and loading."
+        )
+
+    def test_bria_image_output_shape(self):
+        pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device)
+        inputs = self.get_dummy_inputs(torch_device)
+
+        height_width_pairs = [(16, 16), (32, 32), (64, 64)]
+        for height, width in height_width_pairs:
+            expected_height = height - height % (pipe.vae_scale_factor * 2)
+            expected_width = width - width % (pipe.vae_scale_factor * 2)
+
+            inputs.update({"height": height, "width": width})
+            image = pipe(**inputs).images[0]
+            output_height, output_width, _ = image.shape
+            assert (output_height, output_width) == (expected_height, expected_width)
+
+    def test_to_dtype(self):
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.set_progress_bar_config(disable=None)
+
+        model_dtypes = [component.dtype for component in components.values() if hasattr(component, "dtype")]
+        self.assertTrue([dtype == torch.float32 for dtype in model_dtypes] == [True, True, True])
+
+    def test_torch_dtype_dict(self):
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            pipe.save_pretrained(tmpdirname)
+            torch_dtype_dict = {"transformer": torch.bfloat16, "default": torch.float16}
+            loaded_pipe = self.pipeline_class.from_pretrained(tmpdirname, torch_dtype=torch_dtype_dict)
+
+            self.assertEqual(loaded_pipe.transformer.dtype, torch.bfloat16)
+            self.assertEqual(loaded_pipe.text_encoder.dtype, torch.float16)
+            self.assertEqual(loaded_pipe.vae.dtype, torch.float16)
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            pipe.save_pretrained(tmpdirname)
+            torch_dtype_dict = {"default": torch.float16}
+            loaded_pipe = self.pipeline_class.from_pretrained(tmpdirname, torch_dtype=torch_dtype_dict)
+
+            self.assertEqual(loaded_pipe.transformer.dtype, torch.float16)
+            self.assertEqual(loaded_pipe.text_encoder.dtype, torch.float16)
+            self.assertEqual(loaded_pipe.vae.dtype, torch.float16)
+
+
+@slow
+@require_torch_gpu
+class BriaPipelineSlowTests(unittest.TestCase):
+    pipeline_class = BriaPipeline
+    repo_id = "briaai/BRIA-3.2"
+
+    def setUp(self):
+        super().setUp()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def tearDown(self):
+        super().tearDown()
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def get_inputs(self, device, seed=0):
+        generator = torch.Generator(device="cpu").manual_seed(seed)
+
+        prompt_embeds = torch.load(
+            hf_hub_download(repo_id="diffusers/test-slices", repo_type="dataset", filename="flux/prompt_embeds.pt")
+        ).to(torch_device)
+
+        return {
+            "prompt_embeds": prompt_embeds,
+            "num_inference_steps": 2,
+            "guidance_scale": 0.0,
+            "max_sequence_length": 256,
+            "output_type": "np",
+            "generator": generator,
+        }
+
+    def test_bria_inference_bf16(self):
+        pipe = self.pipeline_class.from_pretrained(
+            self.repo_id, torch_dtype=torch.bfloat16, text_encoder=None, tokenizer=None
+        )
+        pipe.to(torch_device)
+
+        inputs = self.get_inputs(torch_device)
+
+        image = pipe(**inputs).images[0]
+        image_slice = image[0, :10, :10].flatten()
+
+        expected_slice = np.array(
+            [
+                0.59729785,
+                0.6153719,
+                0.595112,
+                0.5884763,
+                0.59366125,
+                0.5795311,
+                0.58325,
+                0.58449626,
+                0.57737637,
+                0.58432233,
+                0.5867875,
+                0.57824117,
+                0.5819089,
+                0.5830988,
+                0.57730293,
+                0.57647324,
+                0.5769151,
+                0.57312685,
+                0.57926565,
+                0.5823928,
+                0.57783926,
+                0.57162863,
+                0.575649,
+                0.5745547,
+                0.5740556,
+                0.5799735,
+                0.57799566,
+                0.5715559,
+                0.5771242,
+                0.5773058,
+            ],
+            dtype=np.float32,
+        )
+        max_diff = numpy_cosine_similarity_distance(expected_slice, image_slice)
+        self.assertLess(max_diff, 1e-4, f"Image slice is different from expected slice: {max_diff:.4f}")