review fixes

openvinotoolkit · Jan 23, 2025 · b06c567 · b06c567
1 parent 0d1ec0b
commit b06c567
Show file tree

Hide file tree

Showing 6 changed files with 13 additions and 16 deletions.
diff --git a/SUPPORTED_MODELS.md b/SUPPORTED_MODELS.md
@@ -242,7 +242,7 @@ The pipeline can work with other similar topologies produced by `optimum-intel`
       <tr>
       <td><code>Flux</code></td>
       <td>Supported</td>
-      <td>Not supported</td>
+      <td>Supported</td>
       <td>Not supported</td>
       <td>
         <ul>

diff --git a/src/cpp/src/image_generation/flux_pipeline.hpp b/src/cpp/src/image_generation/flux_pipeline.hpp
@@ -429,15 +429,6 @@ class FluxPipeline : public DiffusionPipeline {
         const size_t vae_scale_factor = m_vae->get_vae_scale_factor();
         const auto& transformer_config = m_transformer->get_config();
 
-        // in case of image to image generation_config_value is just ignored and computed based on initial image
-        if (m_pipeline_type == PipelineType::IMAGE_2_IMAGE) {
-            OPENVINO_ASSERT(initial_image, "Initial image is empty for image to image pipeline");
-            ov::Shape shape = initial_image.get_shape();
-            int64_t dim_val = shape[dim_idx];
-
-            generation_config_value = dim_val - (dim_val % vae_scale_factor);
-        }
-
         if (generation_config_value < 0)
             generation_config_value = transformer_config.m_default_sample_size * vae_scale_factor;
     }
@@ -451,6 +442,9 @@ class FluxPipeline : public DiffusionPipeline {
 
         m_generation_config = ImageGenerationConfig();
 
+        m_generation_config.height = transformer_config.m_default_sample_size * vae_scale_factor;
+        m_generation_config.width = transformer_config.m_default_sample_size * vae_scale_factor;
+
         if (class_name == "FluxPipeline" || class_name == "FluxImg2ImgPipeline" || class_name == "FluxInpaintPipeline" ) {
             if (m_pipeline_type == PipelineType::TEXT_2_IMAGE) {
                 m_generation_config.guidance_scale = 3.5f;
@@ -460,8 +454,6 @@ class FluxPipeline : public DiffusionPipeline {
                 m_generation_config.guidance_scale = 7.0f;
                 m_generation_config.num_inference_steps = 28;
                 m_generation_config.strength = 0.6f;
-                m_generation_config.height = 1024;
-                m_generation_config.width = 1024;
             }
             m_generation_config.max_sequence_length = 512;
         } else {

diff --git a/src/cpp/src/image_generation/schedulers/flow_match_euler_discrete.hpp b/src/cpp/src/image_generation/schedulers/flow_match_euler_discrete.hpp
@@ -50,8 +50,6 @@ class FlowMatchEulerDiscreteScheduler : public IScheduler {
 
     size_t get_begin_index() override;
 
-
-
 private:
     Config m_config;
 

diff --git a/src/python/openvino_genai/py_openvino_genai.pyi b/src/python/openvino_genai/py_openvino_genai.pyi
@@ -774,6 +774,9 @@ class Image2ImagePipeline:
     This class is used for generation with image-to-image models.
     """
     @staticmethod
+    def flux(scheduler: Scheduler, clip_text_model: CLIPTextModel, t5_encoder_model: T5EncoderModel, transformer: FluxTransformer2DModel, vae: AutoencoderKL) -> Image2ImagePipeline:
+        ...
+    @staticmethod
     def latent_consistency_model(scheduler: Scheduler, clip_text_model: CLIPTextModel, unet: UNet2DConditionModel, vae: AutoencoderKL) -> Image2ImagePipeline:
         ...
     @staticmethod

diff --git a/src/python/py_image_generation_pipelines.cpp b/src/python/py_image_generation_pipelines.cpp
@@ -330,6 +330,7 @@ void init_image_generation_pipelines(py::module_& m) {
         .def_static("stable_diffusion", &ov::genai::Image2ImagePipeline::stable_diffusion, py::arg("scheduler"), py::arg("clip_text_model"), py::arg("unet"), py::arg("vae"))
         .def_static("latent_consistency_model", &ov::genai::Image2ImagePipeline::latent_consistency_model, py::arg("scheduler"), py::arg("clip_text_model"), py::arg("unet"), py::arg("vae"))
         .def_static("stable_diffusion_xl", &ov::genai::Image2ImagePipeline::stable_diffusion_xl, py::arg("scheduler"), py::arg("clip_text_model"), py::arg("clip_text_model_with_projection"), py::arg("unet"), py::arg("vae"))
+        .def_static("flux", &ov::genai::Image2ImagePipeline::flux, py::arg("scheduler"), py::arg("clip_text_model"), py::arg("t5_encoder_model"), py::arg("transformer"), py::arg("vae"))
         .def(
             "compile",
             [](ov::genai::Image2ImagePipeline& pipe,

diff --git a/tools/who_what_benchmark/tests/test_cli_image.py b/tools/who_what_benchmark/tests/test_cli_image.py
@@ -103,8 +103,11 @@ def test_image_model_types(model_id, model_type, backend):
                             ])),
 )
 def test_image_model_genai(model_id, model_type):
-    if ("flux" in model_id or "stable-diffusion-3" in model_id) and model_type != "text-to-image":
-        pytest.skip(reason="FLUX or SD3 are supported as text to image only")
+    if ("stable-diffusion-3" in model_id) and model_type != "text-to-image":
+        pytest.skip(reason="SD3 is supported as text to image only")
+
+    if ("flux" in model_id) and model_type != "image-inpainting":
+        pytest.skip(reason="FLUX is not yet supported as image inpainting")
 
     with tempfile.TemporaryDirectory() as temp_dir:
         GT_FILE = os.path.join(temp_dir, "gt.csv")
Original file line number	Diff line number	Diff line change
Expand Up		@@ -50,8 +50,6 @@ class FlowMatchEulerDiscreteScheduler : public IScheduler {

		size_t get_begin_index() override;



		private:
		Config m_config;

Expand Down