feat(diffusers): add support for wan2.2

Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

backend/backend.proto

@@ -312,15 +312,17 @@ message GenerateImageRequest {
 
 message GenerateVideoRequest {
   string prompt = 1;
-  string start_image = 2;  // Path or base64 encoded image for the start frame
-  string end_image = 3;    // Path or base64 encoded image for the end frame
-  int32 width = 4;
-  int32 height = 5;
-  int32 num_frames = 6;    // Number of frames to generate
-  int32 fps = 7;          // Frames per second
-  int32 seed = 8;
-  float cfg_scale = 9;    // Classifier-free guidance scale
-  string dst = 10;        // Output path for the generated video
+  string negative_prompt = 2;  // Negative prompt for video generation
+  string start_image = 3;  // Path or base64 encoded image for the start frame
+  string end_image = 4;    // Path or base64 encoded image for the end frame
+  int32 width = 5;
+  int32 height = 6;
+  int32 num_frames = 7;    // Number of frames to generate
+  int32 fps = 8;          // Frames per second
+  int32 seed = 9;
+  float cfg_scale = 10;    // Classifier-free guidance scale
+  int32 step = 11;         // Number of inference steps
+  string dst = 12;        // Output path for the generated video
 }
 
 message TTSRequest {

backend/python/diffusers/backend.py

@@ -18,7 +18,7 @@ import backend_pb2_grpc
 import grpc
 
 from diffusers import SanaPipeline, StableDiffusion3Pipeline, StableDiffusionXLPipeline, StableDiffusionDepth2ImgPipeline, DPMSolverMultistepScheduler, StableDiffusionPipeline, DiffusionPipeline, \
-    EulerAncestralDiscreteScheduler, FluxPipeline, FluxTransformer2DModel, QwenImageEditPipeline
+    EulerAncestralDiscreteScheduler, FluxPipeline, FluxTransformer2DModel, QwenImageEditPipeline, AutoencoderKLWan, WanPipeline, WanImageToVideoPipeline
 from diffusers import StableDiffusionImg2ImgPipeline, AutoPipelineForText2Image, ControlNetModel, StableVideoDiffusionPipeline, Lumina2Text2ImgPipeline
 from diffusers.pipelines.stable_diffusion import safety_checker
 from diffusers.utils import load_image, export_to_video
@@ -334,6 +334,32 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
                     torch_dtype=torch.bfloat16)
                 self.pipe.vae.to(torch.bfloat16)
                 self.pipe.text_encoder.to(torch.bfloat16)
+            elif request.PipelineType == "WanPipeline":
+                # WAN2.2 pipeline requires special VAE handling
+                vae = AutoencoderKLWan.from_pretrained(
+                    request.Model, 
+                    subfolder="vae", 
+                    torch_dtype=torch.float32
+                )
+                self.pipe = WanPipeline.from_pretrained(
+                    request.Model,
+                    vae=vae,
+                    torch_dtype=torchType
+                )
+                self.txt2vid = True  # WAN2.2 is a text-to-video pipeline
+            elif request.PipelineType == "WanImageToVideoPipeline":
+                # WAN2.2 image-to-video pipeline
+                vae = AutoencoderKLWan.from_pretrained(
+                    request.Model, 
+                    subfolder="vae", 
+                    torch_dtype=torch.float32
+                )
+                self.pipe = WanImageToVideoPipeline.from_pretrained(
+                    request.Model,
+                    vae=vae,
+                    torch_dtype=torchType
+                )
+                self.img2vid = True  # WAN2.2 image-to-video pipeline
 
             if CLIPSKIP and request.CLIPSkip != 0:
                 self.clip_skip = request.CLIPSkip
@@ -575,6 +601,96 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
 
         return backend_pb2.Result(message="Media generated", success=True)
 
+    def GenerateVideo(self, request, context):
+        try:
+            prompt = request.prompt
+            if not prompt:
+                return backend_pb2.Result(success=False, message="No prompt provided for video generation")
+
+            # Set default values from request or use defaults
+            num_frames = request.num_frames if request.num_frames > 0 else 81
+            fps = request.fps if request.fps > 0 else 16
+            cfg_scale = request.cfg_scale if request.cfg_scale > 0 else 4.0
+            num_inference_steps = request.step if request.step > 0 else 40
+            
+            # Prepare generation parameters
+            kwargs = {
+                "prompt": prompt,
+                "negative_prompt": request.negative_prompt if request.negative_prompt else "",
+                "height": request.height if request.height > 0 else 720,
+                "width": request.width if request.width > 0 else 1280,
+                "num_frames": num_frames,
+                "guidance_scale": cfg_scale,
+                "num_inference_steps": num_inference_steps,
+            }
+
+            # Add custom options from self.options (including guidance_scale_2 if specified)
+            kwargs.update(self.options)
+
+            # Set seed if provided
+            if request.seed > 0:
+                kwargs["generator"] = torch.Generator(device=self.device).manual_seed(request.seed)
+
+            # Handle start and end images for video generation
+            if request.start_image:
+                kwargs["start_image"] = load_image(request.start_image)
+            if request.end_image:
+                kwargs["end_image"] = load_image(request.end_image)
+
+            print(f"Generating video with {kwargs=}", file=sys.stderr)
+
+            # Generate video frames based on pipeline type
+            if self.PipelineType == "WanPipeline":
+                # WAN2.2 text-to-video generation
+                output = self.pipe(**kwargs)
+                frames = output.frames[0]  # WAN2.2 returns frames in this format
+            elif self.PipelineType == "WanImageToVideoPipeline":
+                # WAN2.2 image-to-video generation
+                if request.start_image:
+                    # Load and resize the input image according to WAN2.2 requirements
+                    image = load_image(request.start_image)
+                    # Use request dimensions or defaults, but respect WAN2.2 constraints
+                    request_height = request.height if request.height > 0 else 480
+                    request_width = request.width if request.width > 0 else 832
+                    max_area = request_height * request_width
+                    aspect_ratio = image.height / image.width
+                    mod_value = self.pipe.vae_scale_factor_spatial * self.pipe.transformer.config.patch_size[1]
+                    height = round((max_area * aspect_ratio) ** 0.5 / mod_value) * mod_value
+                    width = round((max_area / aspect_ratio) ** 0.5 / mod_value) * mod_value
+                    image = image.resize((width, height))
+                    kwargs["image"] = image
+                    kwargs["height"] = height
+                    kwargs["width"] = width
+                
+                output = self.pipe(**kwargs)
+                frames = output.frames[0]
+            elif self.img2vid:
+                # Generic image-to-video generation
+                if request.start_image:
+                    image = load_image(request.start_image)
+                    image = image.resize((request.width if request.width > 0 else 1024, 
+                                       request.height if request.height > 0 else 576))
+                    kwargs["image"] = image
+                
+                output = self.pipe(**kwargs)
+                frames = output.frames[0]
+            elif self.txt2vid:
+                # Generic text-to-video generation
+                output = self.pipe(**kwargs)
+                frames = output.frames[0]
+            else:
+                return backend_pb2.Result(success=False, message=f"Pipeline {self.PipelineType} does not support video generation")
+
+            # Export video
+            export_to_video(frames, request.dst, fps=fps)
+            
+            return backend_pb2.Result(message="Video generated successfully", success=True)
+
+        except Exception as err:
+            print(f"Error generating video: {err}", file=sys.stderr)
+            traceback.print_exc()
+            return backend_pb2.Result(success=False, message=f"Error generating video: {err}")
+
 
 def serve(address):
     server = grpc.server(futures.ThreadPoolExecutor(max_workers=MAX_WORKERS),

core/backend/video.go

@@ -7,7 +7,7 @@ import (
 	model "github.com/mudler/LocalAI/pkg/model"
 )
 
-func VideoGeneration(height, width int32, prompt, startImage, endImage, dst string, loader *model.ModelLoader, modelConfig config.ModelConfig, appConfig *config.ApplicationConfig) (func() error, error) {
+func VideoGeneration(height, width int32, prompt, negativePrompt, startImage, endImage, dst string, numFrames, fps, seed int32, cfgScale float32, step int32, loader *model.ModelLoader, modelConfig config.ModelConfig, appConfig *config.ApplicationConfig) (func() error, error) {
 
 	opts := ModelOptions(modelConfig, appConfig)
 	inferenceModel, err := loader.Load(
@@ -22,12 +22,18 @@ func VideoGeneration(height, width int32, prompt, startImage, endImage, dst stri
 		_, err := inferenceModel.GenerateVideo(
 			appConfig.Context,
 			&proto.GenerateVideoRequest{
-				Height:     height,
-				Width:      width,
-				Prompt:     prompt,
-				StartImage: startImage,
-				EndImage:   endImage,
-				Dst:        dst,
+				Height:         height,
+				Width:          width,
+				Prompt:         prompt,
+				NegativePrompt: negativePrompt,
+				StartImage:     startImage,
+				EndImage:       endImage,
+				NumFrames:      numFrames,
+				Fps:            fps,
+				Seed:           seed,
+				CfgScale:       cfgScale,
+				Step:           step,
+				Dst:            dst,
 			})
 		return err
 	}

core/http/endpoints/localai/video.go

@@ -61,7 +61,7 @@ func downloadFile(url string) (string, error) {
 */
 // VideoEndpoint
 // @Summary Creates a video given a prompt.
-// @Param request body schema.OpenAIRequest true "query params"
+// @Param request body schema.VideoRequest true "query params"
 // @Success 200 {object} schema.OpenAIResponse "Response"
 // @Router /video [post]
 func VideoEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, appConfig *config.ApplicationConfig) func(c *fiber.Ctx) error {
@@ -166,7 +166,23 @@ func VideoEndpoint(cl *config.ModelConfigLoader, ml *model.ModelLoader, appConfi
 
 		baseURL := c.BaseURL()
 
-		fn, err := backend.VideoGeneration(height, width, input.Prompt, src, input.EndImage, output, ml, *config, appConfig)
+		fn, err := backend.VideoGeneration(
+			height,
+			width,
+			input.Prompt,
+			input.NegativePrompt,
+			src,
+			input.EndImage,
+			output,
+			input.NumFrames,
+			input.FPS,
+			input.Seed,
+			input.CFGScale,
+			input.Step,
+			ml,
+			*config,
+			appConfig,
+		)
 		if err != nil {
 			return err
 		}

core/schema/localai.go

@@ -28,6 +28,7 @@ type GalleryResponse struct {
 type VideoRequest struct {
 	BasicModelRequest
 	Prompt         string  `json:"prompt" yaml:"prompt"`
+	NegativePrompt string  `json:"negative_prompt" yaml:"negative_prompt"`
 	StartImage     string  `json:"start_image" yaml:"start_image"`
 	EndImage       string  `json:"end_image" yaml:"end_image"`
 	Width          int32   `json:"width" yaml:"width"`
@@ -36,6 +37,7 @@ type VideoRequest struct {
 	FPS            int32   `json:"fps" yaml:"fps"`
 	Seed           int32   `json:"seed" yaml:"seed"`
 	CFGScale       float32 `json:"cfg_scale" yaml:"cfg_scale"`
+	Step           int32   `json:"step" yaml:"step"`
 	ResponseFormat string  `json:"response_format" yaml:"response_format"`
 }