community-efforts/image_preferences/01_synthetic_data_generation_images.py [23:296]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - class InferenceEndpointsImageLLM(InferenceEndpointsLLM): @validate_call async def agenerate( self, input: Dict[str, Any], negative_prompt: Optional[str] = None, height: Optional[float] = None, width: Optional[float] = None, num_inference_steps: Optional[float] = None, guidance_scale: Optional[float] = None, ) -> List[Dict[str, Any]]: prompt = input.get("prompt") image = await self._aclient.text_to_image( prompt, negative_prompt=negative_prompt, height=int(height) if height else None, width=int(width) if width else None, num_inference_steps=int(num_inference_steps) if num_inference_steps else None, guidance_scale=float(guidance_scale) if guidance_scale else None, seed=random.randint(0, 1000000), ) buffered = BytesIO() image.save(buffered, format="JPEG") img_str = base64.b64encode(buffered.getvalue()).decode() return [{"image": img_str}] class ImageGeneration(Task): @property def inputs(self) -> List[str]: return ["prompt"] @property def outputs(self) -> List[str]: return ["image", "model_name"] def format_input(self, input: Dict[str, Any]) -> Dict[str, str]: return {"prompt": input["prompt"]} def format_output( self, output: Dict[str, Any], input: Dict[str, Any] ) -> Dict[str, Any]: image_str = output.get("image") image = None if image_str: image_bytes = base64.b64decode(image_str) image = Image.open(BytesIO(image_bytes)) return {"image": image, "model_name": self.llm.model_name} def process(self, *args: StepInput) -> "StepOutput": inputs = args[0] if args else [] formatted_inputs = self._format_inputs(inputs) outputs = self.llm.generate_outputs( inputs=formatted_inputs, num_generations=self.num_generations, **self.llm.get_generation_kwargs(), ) task_outputs = [] for input, input_outputs in zip(inputs, outputs): formatted_outputs = self._format_outputs(input_outputs, input) for formatted_output in formatted_outputs: if "image" in formatted_output and formatted_output["image"]: # use prompt as filename prompt_hash = hashlib.md5(input["prompt"].encode()).hexdigest() self.save_artifact( name="images", write_function=lambda path: formatted_output["image"].save( path / f"{prompt_hash}.jpeg" ), metadata={"type": "image", "library": "diffusers"}, ) formatted_output["image"] = { "path": f"artifacts/{self.name}/images/{prompt_hash}.jpeg" } task_output = { **input, **formatted_output, "model_name": self.llm.model_name, } task_outputs.append(task_output) yield task_outputs ## Let's determine the categories and subcategories for the image generation task # https://huggingface.co/spaces/google/sdxl/blob/main/app.py#L55 categories = { # included "Cinematic": [ # included "emotional", "harmonious", "vignette", "highly detailed", "high budget", "bokeh", "cinemascope", "moody", "epic", "gorgeous", "film grain", "grainy", ], # included "Photographic": [ # included "film", "bokeh", "professional", "4k", "highly detailed", ## not included "Landscape", "Portrait", "Macro", "Portra", "Gold", "ColorPlus", "Ektar", "Superia", "C200", "CineStill", "CineStill 50D", "CineStill 800T", "Tri-X", "HP5", "Delta", "T-Max", "Fomapan", "StreetPan", "Provia", "Ektachrome", "Velvia", ], # included "Anime": [ # included "anime style", "key visual", "vibrant", "studio anime", "highly detailed", ], # included "Manga": [ # included "vibrant", "high-energy", "detailed", "iconic", "Japanese comic style", ], # included "Digital art": [ # included "digital artwork", "illustrative", "painterly", "matte painting", "highly detailed", ], # included "Pixel art": [ # included "low-res", "blocky", "pixel art style", "8-bit graphics", ], # included "Fantasy art": [ # included "magnificent", "celestial", "ethereal", "painterly", "epic", "majestic", "magical", "fantasy art", "cover art", "dreamy", ], # included "Neonpunk": [ # included "cyberpunk", "vaporwave", "neon", "vibes", "vibrant", "stunningly beautiful", "crisp", "detailed", "sleek", "ultramodern", "magenta highlights", "dark purple shadows", "high contrast", "cinematic", "ultra detailed", "intricate", "professional", ], # included "3D Model": [ # included "octane render", "highly detailed", "volumetric", "dramatic lighting", ], # not included "Painting": [ "Oil", "Acrylic", "Watercolor", "Digital", "Mural", "Sketch", "Gouache", "Renaissance", "Baroque", "Romanticism", "Impressionism", "Expressionism", "Cubism", "Surrealism", "Pop Art", "Minimalism", "Realism", "Encaustic", "Tempera", "Fresco", "Ink Wash", "Spray Paint", "Mixed Media", ], # not included "Animation": [ # not included "Animation", "Stop motion", "Claymation", "Pixel Art", "Vector", "Hand-drawn", "Cutout", "Whiteboard", ], # not included "Illustration": [ # not included "Book", "Comics", "Editorial", "Advertising", "Technical", "Fantasy", "Scientific", "Fashion", "Storyboard", "Concept Art", "Manga", "Anime", "Digital", "Vector", "Design", ], } - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - community-efforts/image_preferences/01_synthetic_data_generation_total.py [23:296]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - class InferenceEndpointsImageLLM(InferenceEndpointsLLM): @validate_call async def agenerate( self, input: Dict[str, Any], negative_prompt: Optional[str] = None, height: Optional[float] = None, width: Optional[float] = None, num_inference_steps: Optional[float] = None, guidance_scale: Optional[float] = None, ) -> List[Dict[str, Any]]: prompt = input.get("prompt") image = await self._aclient.text_to_image( prompt, negative_prompt=negative_prompt, height=int(height) if height else None, width=int(width) if width else None, num_inference_steps=int(num_inference_steps) if num_inference_steps else None, guidance_scale=float(guidance_scale) if guidance_scale else None, seed=random.randint(0, 1000000), ) buffered = BytesIO() image.save(buffered, format="JPEG") img_str = base64.b64encode(buffered.getvalue()).decode() return [{"image": img_str}] class ImageGeneration(Task): @property def inputs(self) -> List[str]: return ["prompt"] @property def outputs(self) -> List[str]: return ["image", "model_name"] def format_input(self, input: Dict[str, Any]) -> Dict[str, str]: return {"prompt": input["prompt"]} def format_output( self, output: Dict[str, Any], input: Dict[str, Any] ) -> Dict[str, Any]: image_str = output.get("image") image = None if image_str: image_bytes = base64.b64decode(image_str) image = Image.open(BytesIO(image_bytes)) return {"image": image, "model_name": self.llm.model_name} def process(self, *args: StepInput) -> "StepOutput": inputs = args[0] if args else [] formatted_inputs = self._format_inputs(inputs) outputs = self.llm.generate_outputs( inputs=formatted_inputs, num_generations=self.num_generations, **self.llm.get_generation_kwargs(), ) task_outputs = [] for input, input_outputs in zip(inputs, outputs): formatted_outputs = self._format_outputs(input_outputs, input) for formatted_output in formatted_outputs: if "image" in formatted_output and formatted_output["image"]: # use prompt as filename prompt_hash = hashlib.md5(input["prompt"].encode()).hexdigest() self.save_artifact( name="images", write_function=lambda path: formatted_output["image"].save( path / f"{prompt_hash}.jpeg" ), metadata={"type": "image", "library": "diffusers"}, ) formatted_output["image"] = { "path": f"artifacts/{self.name}/images/{prompt_hash}.jpeg" } task_output = { **input, **formatted_output, "model_name": self.llm.model_name, } task_outputs.append(task_output) yield task_outputs ## Let's determine the categories and subcategories for the image generation task # https://huggingface.co/spaces/google/sdxl/blob/main/app.py#L55 categories = { # included "Cinematic": [ # included "emotional", "harmonious", "vignette", "highly detailed", "high budget", "bokeh", "cinemascope", "moody", "epic", "gorgeous", "film grain", "grainy", ], # included "Photographic": [ # included "film", "bokeh", "professional", "4k", "highly detailed", ## not included "Landscape", "Portrait", "Macro", "Portra", "Gold", "ColorPlus", "Ektar", "Superia", "C200", "CineStill", "CineStill 50D", "CineStill 800T", "Tri-X", "HP5", "Delta", "T-Max", "Fomapan", "StreetPan", "Provia", "Ektachrome", "Velvia", ], # included "Anime": [ # included "anime style", "key visual", "vibrant", "studio anime", "highly detailed", ], # included "Manga": [ # included "vibrant", "high-energy", "detailed", "iconic", "Japanese comic style", ], # included "Digital art": [ # included "digital artwork", "illustrative", "painterly", "matte painting", "highly detailed", ], # included "Pixel art": [ # included "low-res", "blocky", "pixel art style", "8-bit graphics", ], # included "Fantasy art": [ # included "magnificent", "celestial", "ethereal", "painterly", "epic", "majestic", "magical", "fantasy art", "cover art", "dreamy", ], # included "Neonpunk": [ # included "cyberpunk", "vaporwave", "neon", "vibes", "vibrant", "stunningly beautiful", "crisp", "detailed", "sleek", "ultramodern", "magenta highlights", "dark purple shadows", "high contrast", "cinematic", "ultra detailed", "intricate", "professional", ], # included "3D Model": [ # included "octane render", "highly detailed", "volumetric", "dramatic lighting", ], # not included "Painting": [ "Oil", "Acrylic", "Watercolor", "Digital", "Mural", "Sketch", "Gouache", "Renaissance", "Baroque", "Romanticism", "Impressionism", "Expressionism", "Cubism", "Surrealism", "Pop Art", "Minimalism", "Realism", "Encaustic", "Tempera", "Fresco", "Ink Wash", "Spray Paint", "Mixed Media", ], # not included "Animation": [ # not included "Animation", "Stop motion", "Claymation", "Pixel Art", "Vector", "Hand-drawn", "Cutout", "Whiteboard", ], # not included "Illustration": [ # not included "Book", "Comics", "Editorial", "Advertising", "Technical", "Fantasy", "Scientific", "Fashion", "Storyboard", "Concept Art", "Manga", "Anime", "Digital", "Vector", "Design", ], } - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -