from pathlib import Path import gradio as gr import torch from finetuning import FineTunedModel from StableDiffuser import StableDiffuser from tqdm import tqdm model_map = { 'Car' : 'models/car.pt', 'Van Gogh' : 'models/vangogh.pt', } class Demo: def __init__(self) -> None: self.training = False self.generating = False self.nsteps = 50 self.diffuser = StableDiffuser(scheduler='DDIM', seed=42).to('cuda') self.finetuner = None with gr.Blocks() as demo: self.layout() self.switch_model(self.model_dropdown.value) self.finetuner = self.finetuner.eval().half() self.diffuser = self.diffuser.eval().half() demo.queue(concurrency_count=2).launch() def disable(self): return [gr.update(interactive=False), gr.update(interactive=False)] def switch_model(self, model_name): if not model_name: return model_path = model_map[model_name] checkpoint = torch.load(model_path) self.finetuner = FineTunedModel.from_checkpoint(self.diffuser, checkpoint) torch.cuda.empty_cache() def layout(self): with gr.Row(): with gr.Tab("Test") as inference_column: with gr.Row(): self.explain_infr = gr.Markdown(interactive=False, value='This is a demo of [Erasing Concepts from Stable Diffusion](https://erasing.baulab.info/). To try out a model where a concept has been erased, select a model and enter any prompt. For example, if you select the model "Van Gogh" you can generate images for the prompt "A portrait in the style of Van Gogh" and compare the erased and unerased models. We have also provided models with "cars" erased, and with "nudity" erased. You can also train and run your own custom model with a concept erased.') with gr.Row(): with gr.Column(scale=1): self.prompt_input_infr = gr.Text( placeholder="Enter prompt...", label="Prompt", info="Prompt to generate" ) with gr.Row(): self.model_dropdown = gr.Dropdown( label="ESD Model", choices=['Van Gogh', 'Car'], value='Van Gogh', interactive=True ) self.seed_infr = gr.Number( label="Seed", value=42 ) with gr.Column(scale=2): self.infr_button = gr.Button( value="Generate", interactive=True ) with gr.Row(): self.image_new = gr.Image( label="ESD", interactive=False ) self.image_orig = gr.Image( label="SD", interactive=False ) with gr.Tab("Train") as training_column: with gr.Row(): self.explain_train= gr.Markdown(interactive=False, value='In this part you can erase any concept from Stable Diffusion. Enter a prompt for the concept or style you want to erase, and select ESD-x if you want to focus erasure on prompts that mention the concept explicitly, or ESD-u if you want to erase the concept even for prompts that do not mention the concept. With default settings, it takes about 20 minutes to fine-tune the model; then you can try inference above or download the weights. The training code used here is slightly different than the code tested in the original paper. Code and details are at [github link](https://github.com/rohitgandikota/erasing).') with gr.Row(): with gr.Column(scale=3): self.prompt_input = gr.Text( placeholder="Enter prompt...", label="Prompt to Erase", info="Prompt corresponding to concept to erase" ) self.train_method_input = gr.Dropdown( choices=['ESD-x', 'ESD-u', 'ESD-self'], value='ESD-x', label='Train Method', info='Method of training' ) self.neg_guidance_input = gr.Number( value=1, label="Negative Guidance", info='Guidance of negative training used to train' ) self.iterations_input = gr.Number( value=150, precision=0, label="Iterations", info='iterations used to train' ) self.lr_input = gr.Number( value=1e-5, label="Learning Rate", info='Learning rate used to train' ) with gr.Column(scale=1): self.train_button = gr.Button( value="Train", ) self.download = gr.Files() self.model_dropdown.change(self.switch_model, inputs=[self.model_dropdown]) self.infr_button.click(self.inference, inputs = [ self.prompt_input_infr, self.seed_infr ], outputs=[ self.image_new, self.image_orig ] ) self.train_button.click(self.disable, outputs=[self.train_button, self.infr_button] ) self.train_button.click(self.train, inputs = [ self.prompt_input, self.train_method_input, self.neg_guidance_input, self.iterations_input, self.lr_input ], outputs=[self.train_button, self.infr_button, self.download, self.model_dropdown] ) def train(self, prompt, train_method, neg_guidance, iterations, lr, pbar = gr.Progress(track_tqdm=True)): if self.training: return [None, None, None] else: self.training = True del self.finetuner torch.cuda.empty_cache() self.diffuser = self.diffuser.train().float() if train_method == 'ESD-x': modules = ".*attn2$" frozen = [] elif train_method == 'ESD-u': modules = "unet$" frozen = [".*attn2$", "unet.time_embedding$", "unet.conv_out$"] elif train_method == 'ESD-self': modules = ".*attn1$" frozen = [] finetuner = FineTunedModel(self.diffuser, modules, frozen_modules=frozen) optimizer = torch.optim.Adam(finetuner.parameters(), lr=lr) criteria = torch.nn.MSELoss() pbar = tqdm(range(iterations)) with torch.no_grad(): neutral_text_embeddings = self.diffuser.get_text_embeddings([''],n_imgs=1) positive_text_embeddings = self.diffuser.get_text_embeddings([prompt],n_imgs=1) for i in pbar: with torch.no_grad(): self.diffuser.set_scheduler_timesteps(self.nsteps) optimizer.zero_grad() iteration = torch.randint(1, self.nsteps - 1, (1,)).item() latents = self.diffuser.get_initial_latents(1, 512, 1) with finetuner: latents_steps, _ = self.diffuser.diffusion( latents, positive_text_embeddings, start_iteration=0, end_iteration=iteration, guidance_scale=3, show_progress=False ) self.diffuser.set_scheduler_timesteps(1000) iteration = int(iteration / self.nsteps * 1000) positive_latents = self.diffuser.predict_noise(iteration, latents_steps[0], positive_text_embeddings, guidance_scale=1) neutral_latents = self.diffuser.predict_noise(iteration, latents_steps[0], neutral_text_embeddings, guidance_scale=1) with finetuner: negative_latents = self.diffuser.predict_noise(iteration, latents_steps[0], positive_text_embeddings, guidance_scale=1) positive_latents.requires_grad = False neutral_latents.requires_grad = False loss = criteria(negative_latents, neutral_latents - (neg_guidance*(positive_latents - neutral_latents))) #loss = criteria(e_n, e_0) works the best try 5000 epochs loss.backward() optimizer.step() ft_path = f"{prompt.lower().replace(' ', '')}.pt" torch.save(finetuner.state_dict(), ft_path) self.finetuner = finetuner.eval().half() self.diffuser = self.diffuser.eval().half() torch.cuda.empty_cache() self.training = False model_map['Custom'] = ft_path return [gr.update(interactive=True), gr.update(interactive=True), ft_path, gr.Dropdown.update(choices=list(model_map.keys()), value='Custom')] def inference(self, prompt, seed, pbar = gr.Progress(track_tqdm=True)): if self.generating: return [None, None] else: self.generating = True self.diffuser._seed = seed or 42 images = self.diffuser( prompt, n_steps=50, reseed=True ) orig_image = images[0][0] torch.cuda.empty_cache() with self.finetuner: images = self.diffuser( prompt, n_steps=50, reseed=True ) edited_image = images[0][0] self.generating = False torch.cuda.empty_cache() return edited_image, orig_image demo = Demo()