import os import gradio as gr import torch from diffusers import StableDiffusionPipeline from PIL import Image import ImageReward as RM # initialize model_id = "runwayml/stable-diffusion-v1-5" pipe = StableDiffusionPipeline.from_pretrained( model_id, torch_dtype=torch.float16, ) model = RM.load("ImageReward-v1.0") images_in_gallery = [] rewards_in_gallery = [] # event functions def generate_images( prompt, magic_words, num, height, width, num_inference_steps, guidance_scale ): global images_in_gallery, rewards_in_gallery if magic_words is not None: prompt += ", ".join(magic_words) images_in_gallery = pipe( prompt, height=height, width=width, num_inference_steps=num_inference_steps, guidance_scale=guidance_scale, num_images_per_prompt=num, ).images rewards_in_gallery = [None] * len(images_in_gallery) return list(zip(images_in_gallery, rewards_in_gallery)) def score_and_rank(prompt): global rewards_in_gallery, images_in_gallery num_not_scored = rewards_in_gallery.count(None) if num_not_scored > 0: images_to_score = images_in_gallery[-num_not_scored:] with torch.no_grad(): ranks, rewards = model.inference_rank(prompt, images_to_score) if not isinstance(rewards, list): rewards = [rewards] rewards_in_gallery = rewards_in_gallery[:-num_not_scored] + rewards outputs = sorted( zip(images_in_gallery, rewards_in_gallery), key=lambda x: x[1], reverse=True ) images_in_gallery = [image for image, _ in outputs] rewards_in_gallery = [reward for _, reward in outputs] return outputs, [ [idx + 1, reward] for idx, reward in enumerate(rewards_in_gallery) ] else: return list(zip(images_in_gallery, rewards_in_gallery)), [ [idx + 1, reward] for idx, reward in enumerate(rewards_in_gallery) ] def upload_images_to_gallery(uploaded_image_files): global images_in_gallery, rewards_in_gallery uploaded_image_file_paths = [file.name for file in uploaded_image_files] uploaded_images = [Image.open(path) for path in uploaded_image_file_paths] for path in uploaded_image_file_paths: os.remove(path) images_in_gallery = images_in_gallery + uploaded_images rewards_in_gallery = rewards_in_gallery + [None] * len(uploaded_images) return list(zip(images_in_gallery, rewards_in_gallery)) def clear_images(): global images_in_gallery, rewards_in_gallery images_in_gallery = [] rewards_in_gallery = [] return None if __name__ == "__main__": # UI with gr.Blocks( theme=gr.themes.Monochrome(), css=r".caption-label { color: black; }", ) as demo: gr.HTML( """

ImageReward Demo

GitHub Repo • 🤗 HF Repo • 🐦 Twitter • 📃 Paper


ImageReward is the first general-purpose text-to-image human preference RM, which is trained on in total 137k pairs of expert comparisons!

The calculation of ImageRewards is based on both the prompt and images.

""" ) with gr.Row(): with gr.Column(): gr.HTML( """

Try ImageReward with only 2 steps:

  1. Click the "Generate" button in the middle of the bottom.
  2. Click the "Score&Rank" button below the gallery.

Finally, just check ImageRewards along with images or on the right of the gallery.


This demo uses runwayml/stable-diffusion-v1-5 as image generation model.

""" ) with gr.Column(): gr.HTML( """

Besides generating images, you can also upload images to score:

  1. Upload images in the bottom right corner.
  2. Change the "Prompt" to correspond to the images.
  3. Click the "Score&Rank" button below the gallery.

For more details about using ImageReward in your own program, check the README.md in our Github Repo.

""" ) with gr.Row(elem_id="outputs_row"): with gr.Column(elem_id="gallery_column", scale=4): gallery = gr.Gallery( label="Images (scored ones sorted)", show_label=False, elem_id="gallery", ).style(columns=4, object_fit="contain", full_width=True) with gr.Column(elem_id="rewards_column"): rewards = gr.Matrix( value=[[None, None]], headers=["Rank", "ImageReward"], datatype="number", ) with gr.Row(): score_and_rank_button = gr.Button("Score&Rank") clear_button = gr.Button("Clear Gallery") with gr.Row().style(equal_height=True): with gr.Column(): prompt = gr.Textbox( label="Prompt", value="A painting of an ocean with clouds and birds, day time, low depth field effect, oil painting, impressionism", ) examples = [ "A painting of an ocean with clouds and birds, day time, low depth field effect, oil painting, impressionism", "A painting of a girl walking in a hallway and suddenly finds a giant sunflower on the floor blocking her way", "Coronation of the sun emperor, digital art, illustration,4k resolution,intricate extremely detailed, depth,vivid colors", "Symmetry!! Product render poster vivid colors divine proportion owl,glowing fog intricate,elegant, highly detailed", "A unicorn in a clearing.it has a single shining horn. volumetric light.by emmanuel shiu, harry potter, eragon", "Highly detailed portrait of a woman with long hairs,stephen bliss. unreal engine, fantasy art by greg rutkowski", "Sculpture made of flame,portrait, female,future, torch,fire,harper's bazaar,vogue, fashion magazine, intricate", ] prompt_examples = gr.Examples( examples=examples, label="Prompt Examples", inputs=[prompt], elem_id="prompt_examples", ) with gr.Column(): choices = [ "HDR, UHD, 4K, 8K, 64K", "highly detailed", "studio lighting", "professional", "trending on artstation", "unreal engine", "vivid colors", ] magic_words = gr.CheckboxGroup( choices=choices, value=choices, type="value", label="Magic Words to Append to Prompt", ) num = gr.Slider(1, 16, step=1, label="Number of images", value=8) height = gr.Slider(256, 2048, step=256, label="Height", value=512) width = gr.Slider(256, 2048, step=256, label="Width", value=512) num_inference_steps = gr.Slider( 0, 200, step=10, label="Number of inference steps", value=50 ) guidance_scale = gr.Slider( 0, 25, step=0.1, label="Guidance scale", value=7.5 ) generate_button = gr.Button("Generate") with gr.Column(): gr.Markdown( """ - To clear all uploaded images, click the **"Clear Gallery"** button above. - To clear the upload list and add additional images, click the **`x` in the upper right corner of the uploading window**. - Additional images will be appended to the gallery, instead of replacing the existing ones. """ ) uploaded_image_files = gr.File( file_count="multiple", file_types=["image"], type="file", label="Upload Images", show_label=True, ) generate_button.click( generate_images, [ prompt, magic_words, num, height, width, num_inference_steps, guidance_scale, ], [gallery], ) score_and_rank_button.click(score_and_rank, [prompt], [gallery, rewards]) uploaded_image_files.upload( upload_images_to_gallery, [uploaded_image_files], [gallery] ) clear_button.click(clear_images, None, [gallery]) demo.launch()