Add Flux inpainting and Flux Img2Img (#9135)

---------
Co-authored-by: yiyixuxu <yixu310@gmail.com>

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

@@ -163,3 +163,15 @@ image.save("flux-fp8-dev.png")
 [[autodoc]] FluxPipeline
 	- all
 	- __call__
+
+## FluxImg2ImgPipeline
+
+[[autodoc]] FluxImg2ImgPipeline
+	- all
+	- __call__
+
+## FluxInpaintPipeline
+
+[[autodoc]] FluxInpaintPipeline
+	- all
+	- __call__

src/diffusers/__init__.py

@@ -258,6 +258,8 @@ else:
             "CogVideoXVideoToVideoPipeline",
             "CycleDiffusionPipeline",
             "FluxControlNetPipeline",
+            "FluxImg2ImgPipeline",
+            "FluxInpaintPipeline",
             "FluxPipeline",
             "HunyuanDiTControlNetPipeline",
             "HunyuanDiTPAGPipeline",
@@ -703,6 +705,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             CogVideoXVideoToVideoPipeline,
             CycleDiffusionPipeline,
             FluxControlNetPipeline,
+            FluxImg2ImgPipeline,
+            FluxInpaintPipeline,
             FluxPipeline,
             HunyuanDiTControlNetPipeline,
             HunyuanDiTPAGPipeline,

src/diffusers/pipelines/__init__.py

@@ -124,7 +124,12 @@ else:
         "AnimateDiffSparseControlNetPipeline",
         "AnimateDiffVideoToVideoPipeline",
     ]
-    _import_structure["flux"] = ["FluxPipeline", "FluxControlNetPipeline"]
+    _import_structure["flux"] = [
+        "FluxControlNetPipeline",
+        "FluxImg2ImgPipeline",
+        "FluxInpaintPipeline",
+        "FluxPipeline",
+    ]
     _import_structure["audioldm"] = ["AudioLDMPipeline"]
     _import_structure["audioldm2"] = [
         "AudioLDM2Pipeline",
@@ -494,7 +499,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
             VersatileDiffusionTextToImagePipeline,
             VQDiffusionPipeline,
         )
-        from .flux import FluxControlNetPipeline, FluxPipeline
+        from .flux import FluxControlNetPipeline, FluxImg2ImgPipeline, FluxInpaintPipeline, FluxPipeline
         from .hunyuandit import HunyuanDiTPipeline
         from .i2vgen_xl import I2VGenXLPipeline
         from .kandinsky import (

src/diffusers/pipelines/flux/__init__.py

@@ -24,6 +24,8 @@ except OptionalDependencyNotAvailable:
 else:
     _import_structure["pipeline_flux"] = ["FluxPipeline"]
     _import_structure["pipeline_flux_controlnet"] = ["FluxControlNetPipeline"]
+    _import_structure["pipeline_flux_img2img"] = ["FluxImg2ImgPipeline"]
+    _import_structure["pipeline_flux_inpaint"] = ["FluxInpaintPipeline"]
 if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     try:
         if not (is_transformers_available() and is_torch_available()):
@@ -33,6 +35,8 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
     else:
         from .pipeline_flux import FluxPipeline
         from .pipeline_flux_controlnet import FluxControlNetPipeline
+        from .pipeline_flux_img2img import FluxImg2ImgPipeline
+        from .pipeline_flux_inpaint import FluxInpaintPipeline
 else:
     import sys
 

src/diffusers/utils/dummy_torch_and_transformers_objects.py

@@ -317,6 +317,36 @@ class FluxControlNetPipeline(metaclass=DummyObject):
         requires_backends(cls, ["torch", "transformers"])
 
 
+class FluxImg2ImgPipeline(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 FluxInpaintPipeline(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 FluxPipeline(metaclass=DummyObject):
     _backends = ["torch", "transformers"]
 

tests/pipelines/flux/test_pipeline_flux_img2img.py

+import random
+import unittest
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+
+from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, FluxImg2ImgPipeline, FluxTransformer2DModel
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    floats_tensor,
+    torch_device,
+)
+
+from ..test_pipelines_common import PipelineTesterMixin
+
+
+enable_full_determinism()
+
+
+@unittest.skipIf(torch_device == "mps", "Flux has a float64 operation which is not supported in MPS.")
+class FluxImg2ImgPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
+    pipeline_class = FluxImg2ImgPipeline
+    params = frozenset(["prompt", "height", "width", "guidance_scale", "prompt_embeds", "pooled_prompt_embeds"])
+    batch_params = frozenset(["prompt"])
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        transformer = FluxTransformer2DModel(
+            patch_size=1,
+            in_channels=4,
+            num_layers=1,
+            num_single_layers=1,
+            attention_head_dim=16,
+            num_attention_heads=2,
+            joint_attention_dim=32,
+            pooled_projection_dim=32,
+            axes_dims_rope=[4, 4, 8],
+        )
+        clip_text_encoder_config = CLIPTextConfig(
+            bos_token_id=0,
+            eos_token_id=2,
+            hidden_size=32,
+            intermediate_size=37,
+            layer_norm_eps=1e-05,
+            num_attention_heads=4,
+            num_hidden_layers=5,
+            pad_token_id=1,
+            vocab_size=1000,
+            hidden_act="gelu",
+            projection_dim=32,
+        )
+
+        torch.manual_seed(0)
+        text_encoder = CLIPTextModel(clip_text_encoder_config)
+
+        torch.manual_seed(0)
+        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+        tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        torch.manual_seed(0)
+        vae = AutoencoderKL(
+            sample_size=32,
+            in_channels=3,
+            out_channels=3,
+            block_out_channels=(4,),
+            layers_per_block=1,
+            latent_channels=1,
+            norm_num_groups=1,
+            use_quant_conv=False,
+            use_post_quant_conv=False,
+            shift_factor=0.0609,
+            scaling_factor=1.5035,
+        )
+
+        scheduler = FlowMatchEulerDiscreteScheduler()
+
+        return {
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "text_encoder_2": text_encoder_2,
+            "tokenizer": tokenizer,
+            "tokenizer_2": tokenizer_2,
+            "transformer": transformer,
+            "vae": vae,
+        }
+
+    def get_dummy_inputs(self, device, seed=0):
+        image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
+        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",
+            "image": image,
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 5.0,
+            "height": 8,
+            "width": 8,
+            "max_sequence_length": 48,
+            "strength": 0.8,
+            "output_type": "np",
+        }
+        return inputs
+
+    def test_flux_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_2"] = "a different prompt"
+        output_different_prompts = pipe(**inputs).images[0]
+
+        max_diff = np.abs(output_same_prompt - output_different_prompts).max()
+
+        # Outputs should be different here
+        # For some reasons, they don't show large differences
+        assert max_diff > 1e-6
+
+    def test_flux_prompt_embeds(self):
+        pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device)
+        inputs = self.get_dummy_inputs(torch_device)
+
+        output_with_prompt = pipe(**inputs).images[0]
+
+        inputs = self.get_dummy_inputs(torch_device)
+        prompt = inputs.pop("prompt")
+
+        (prompt_embeds, pooled_prompt_embeds, text_ids) = pipe.encode_prompt(
+            prompt,
+            prompt_2=None,
+            device=torch_device,
+            max_sequence_length=inputs["max_sequence_length"],
+        )
+        output_with_embeds = pipe(
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            **inputs,
+        ).images[0]
+
+        max_diff = np.abs(output_with_prompt - output_with_embeds).max()
+        assert max_diff < 1e-4

tests/pipelines/flux/test_pipeline_flux_inpaint.py

+import random
+import unittest
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer, CLIPTextConfig, CLIPTextModel, CLIPTokenizer, T5EncoderModel
+
+from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, FluxInpaintPipeline, FluxTransformer2DModel
+from diffusers.utils.testing_utils import (
+    enable_full_determinism,
+    floats_tensor,
+    torch_device,
+)
+
+from ..test_pipelines_common import PipelineTesterMixin
+
+
+enable_full_determinism()
+
+
+@unittest.skipIf(torch_device == "mps", "Flux has a float64 operation which is not supported in MPS.")
+class FluxInpaintPipelineFastTests(unittest.TestCase, PipelineTesterMixin):
+    pipeline_class = FluxInpaintPipeline
+    params = frozenset(["prompt", "height", "width", "guidance_scale", "prompt_embeds", "pooled_prompt_embeds"])
+    batch_params = frozenset(["prompt"])
+
+    def get_dummy_components(self):
+        torch.manual_seed(0)
+        transformer = FluxTransformer2DModel(
+            patch_size=1,
+            in_channels=8,
+            num_layers=1,
+            num_single_layers=1,
+            attention_head_dim=16,
+            num_attention_heads=2,
+            joint_attention_dim=32,
+            pooled_projection_dim=32,
+            axes_dims_rope=[4, 4, 8],
+        )
+        clip_text_encoder_config = CLIPTextConfig(
+            bos_token_id=0,
+            eos_token_id=2,
+            hidden_size=32,
+            intermediate_size=37,
+            layer_norm_eps=1e-05,
+            num_attention_heads=4,
+            num_hidden_layers=5,
+            pad_token_id=1,
+            vocab_size=1000,
+            hidden_act="gelu",
+            projection_dim=32,
+        )
+
+        torch.manual_seed(0)
+        text_encoder = CLIPTextModel(clip_text_encoder_config)
+
+        torch.manual_seed(0)
+        text_encoder_2 = T5EncoderModel.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        tokenizer = CLIPTokenizer.from_pretrained("hf-internal-testing/tiny-random-clip")
+        tokenizer_2 = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-t5")
+
+        torch.manual_seed(0)
+        vae = AutoencoderKL(
+            sample_size=32,
+            in_channels=3,
+            out_channels=3,
+            block_out_channels=(4,),
+            layers_per_block=1,
+            latent_channels=2,
+            norm_num_groups=1,
+            use_quant_conv=False,
+            use_post_quant_conv=False,
+            shift_factor=0.0609,
+            scaling_factor=1.5035,
+        )
+
+        scheduler = FlowMatchEulerDiscreteScheduler()
+
+        return {
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "text_encoder_2": text_encoder_2,
+            "tokenizer": tokenizer,
+            "tokenizer_2": tokenizer_2,
+            "transformer": transformer,
+            "vae": vae,
+        }
+
+    def get_dummy_inputs(self, device, seed=0):
+        image = floats_tensor((1, 3, 32, 32), rng=random.Random(seed)).to(device)
+        mask_image = torch.ones((1, 1, 32, 32)).to(device)
+        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",
+            "image": image,
+            "mask_image": mask_image,
+            "generator": generator,
+            "num_inference_steps": 2,
+            "guidance_scale": 5.0,
+            "height": 32,
+            "width": 32,
+            "max_sequence_length": 48,
+            "strength": 0.8,
+            "output_type": "np",
+        }
+        return inputs
+
+    def test_flux_inpaint_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_2"] = "a different prompt"
+        output_different_prompts = pipe(**inputs).images[0]
+
+        max_diff = np.abs(output_same_prompt - output_different_prompts).max()
+
+        # Outputs should be different here
+        # For some reasons, they don't show large differences
+        assert max_diff > 1e-6
+
+    def test_flux_inpaint_prompt_embeds(self):
+        pipe = self.pipeline_class(**self.get_dummy_components()).to(torch_device)
+        inputs = self.get_dummy_inputs(torch_device)
+
+        output_with_prompt = pipe(**inputs).images[0]
+
+        inputs = self.get_dummy_inputs(torch_device)
+        prompt = inputs.pop("prompt")
+
+        (prompt_embeds, pooled_prompt_embeds, text_ids) = pipe.encode_prompt(
+            prompt,
+            prompt_2=None,
+            device=torch_device,
+            max_sequence_length=inputs["max_sequence_length"],
+        )
+        output_with_embeds = pipe(
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            **inputs,
+        ).images[0]
+
+        max_diff = np.abs(output_with_prompt - output_with_embeds).max()
+        assert max_diff < 1e-4