# Copyright 2022 Lunar Ring. All rights reserved. # Written by Johannes Stelzer, email stelzer@lunar-ring.ai twitter @j_stelzer # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. import os, sys import torch torch.backends.cudnn.benchmark = False import numpy as np import warnings warnings.filterwarnings('ignore') import warnings import torch from tqdm.auto import tqdm from PIL import Image import torch from movie_util import MovieSaver, concatenate_movies from typing import Callable, List, Optional, Union from latent_blending import get_time, yml_save, LatentBlending, add_frames_linear_interp, compare_dicts from stable_diffusion_holder import StableDiffusionHolder torch.set_grad_enabled(False) import gradio as gr import copy from dotenv import find_dotenv, load_dotenv #%% class BlendingFrontend(): def __init__(self, sdh=None): if sdh is None: self.use_debug = True else: self.use_debug = False self.lb = LatentBlending(sdh) self.share = True self.num_inference_steps = 30 self.depth_strength = 0.25 self.seed1 = 42 self.seed2 = 420 self.guidance_scale = 4.0 self.guidance_scale_mid_damper = 0.5 self.mid_compression_scaler = 1.2 self.prompt1 = "" self.prompt2 = "" self.negative_prompt = "" self.list_settings = [] self.state_current = {} self.showing_current = True self.branch1_influence = 0.3 self.branch1_max_depth_influence = 0.6 self.branch1_influence_decay = 0.3 self.parental_influence = 0.1 self.parental_max_depth_influence = 1.0 self.parental_influence_decay = 1.0 self.nmb_branches_final = 9 self.nmb_imgs_show = 5 # don't change self.fps = 30 self.duration_video = 10 self.t_compute_max_allowed = 10 self.dict_multi_trans = {} self.dict_multi_trans_include = {} self.multi_trans_currently_shown = [] self.list_fp_imgs_current = [] self.current_timestamp = None self.nmb_trans_stack = 8 if not self.use_debug: self.lb.sdh.num_inference_steps = self.num_inference_steps self.height = self.lb.sdh.height self.width = self.lb.sdh.width else: self.height = 768 self.width = 768 self.init_save_dir() def init_save_dir(self): load_dotenv(find_dotenv(), verbose=False) try: self.dp_out = os.getenv("dp_out") except Exception as e: self.dp_out = "" # make dummy image def save_empty_image(self): self.fp_img_empty = os.path.join(self.dp_out, 'empty.jpg') Image.fromarray(np.zeros((self.height, self.width, 3), dtype=np.uint8)).save(self.fp_img_empty, quality=5) def randomize_seed1(self): seed = np.random.randint(0, 10000000) self.seed1 = int(seed) print(f"randomize_seed1: new seed = {self.seed1}") return seed def randomize_seed2(self): seed = np.random.randint(0, 10000000) self.seed2 = int(seed) print(f"randomize_seed2: new seed = {self.seed2}") return seed def setup_lb(self, list_ui_elem): # Collect latent blending variables self.state_current = self.get_state_dict() self.lb.set_width(list_ui_elem[list_ui_keys.index('width')]) self.lb.set_height(list_ui_elem[list_ui_keys.index('height')]) self.lb.set_prompt1(list_ui_elem[list_ui_keys.index('prompt1')]) self.lb.set_prompt2(list_ui_elem[list_ui_keys.index('prompt2')]) self.lb.set_negative_prompt(list_ui_elem[list_ui_keys.index('negative_prompt')]) self.lb.guidance_scale = list_ui_elem[list_ui_keys.index('guidance_scale')] self.lb.guidance_scale_mid_damper = list_ui_elem[list_ui_keys.index('guidance_scale_mid_damper')] self.lb.t_compute_max_allowed = list_ui_elem[list_ui_keys.index('duration_compute')] self.lb.num_inference_steps = list_ui_elem[list_ui_keys.index('num_inference_steps')] self.lb.sdh.num_inference_steps = list_ui_elem[list_ui_keys.index('num_inference_steps')] self.duration_video = list_ui_elem[list_ui_keys.index('duration_video')] self.lb.seed1 = list_ui_elem[list_ui_keys.index('seed1')] self.lb.seed2 = list_ui_elem[list_ui_keys.index('seed2')] self.lb.branch1_influence = list_ui_elem[list_ui_keys.index('branch1_influence')] self.lb.branch1_max_depth_influence = list_ui_elem[list_ui_keys.index('branch1_max_depth_influence')] self.lb.branch1_influence_decay = list_ui_elem[list_ui_keys.index('branch1_influence_decay')] self.lb.parental_influence = list_ui_elem[list_ui_keys.index('parental_influence')] self.lb.parental_max_depth_influence = list_ui_elem[list_ui_keys.index('parental_max_depth_influence')] self.lb.parental_influence_decay = list_ui_elem[list_ui_keys.index('parental_influence_decay')] def compute_img1(self, *args): list_ui_elem = args self.setup_lb(list_ui_elem) fp_img1 = os.path.join(self.dp_out, f"img1_{get_time('second')}.jpg") img1 = Image.fromarray(self.lb.compute_latents1(return_image=True)) img1.save(fp_img1) self.save_empty_image() return [fp_img1, self.fp_img_empty, self.fp_img_empty, self.fp_img_empty, self.fp_img_empty] def compute_img2(self, *args): list_ui_elem = args self.setup_lb(list_ui_elem) fp_img2 = os.path.join(self.dp_out, f"img2_{get_time('second')}.jpg") img2 = Image.fromarray(self.lb.compute_latents2(return_image=True)) img2.save(fp_img2) return [self.fp_img_empty, self.fp_img_empty, self.fp_img_empty, fp_img2] def compute_transition(self, *args): list_ui_elem = args self.setup_lb(list_ui_elem) print("STARTING DIFFUSION!") if self.use_debug: list_imgs = [(255*np.random.rand(self.height,self.width,3)).astype(np.uint8) for l in range(5)] list_imgs = [Image.fromarray(l) for l in list_imgs] print("DONE! SENDING BACK RESULTS") return list_imgs fixed_seeds = [self.seed1, self.seed2] # Run Latent Blending imgs_transition = self.lb.run_transition( recycle_img1=True, recycle_img2=True, num_inference_steps=self.num_inference_steps, depth_strength=self.depth_strength, fixed_seeds=fixed_seeds ) print(f"Latent Blending pass finished. Resulted in {len(imgs_transition)} images") # Subselect three preview images idx_img_prev = np.round(np.linspace(0, len(imgs_transition)-1, 5)[1:-1]).astype(np.int32) list_imgs_preview = [] for j in idx_img_prev: list_imgs_preview.append(Image.fromarray(imgs_transition[j])) # Save the preview imgs as jpgs on disk so we are not sending umcompressed data around self.current_timestamp = get_time('second') self.list_fp_imgs_current = [] for i in range(len(list_imgs_preview)): fp_img = f"img_preview_{i}_{self.current_timestamp}.jpg" list_imgs_preview[i].save(fp_img) self.list_fp_imgs_current.append(fp_img) # Insert cheap frames for the movie imgs_transition_ext = add_frames_linear_interp(imgs_transition, self.duration_video, self.fps) # Save as movie fp_movie = self.get_fp_movie(self.current_timestamp) if os.path.isfile(fp_movie): os.remove(fp_movie) ms = MovieSaver(fp_movie, fps=self.fps) for img in tqdm(imgs_transition_ext): ms.write_frame(img) ms.finalize() print("DONE SAVING MOVIE! SENDING BACK...") # Assemble Output, updating the preview images and le movie list_return = self.list_fp_imgs_current + [fp_movie] return list_return def get_fp_movie(self, timestamp, is_stacked=False): if not is_stacked: fn = f"movie_{timestamp}.mp4" else: fn = f"movie_stacked_{timestamp}.mp4" fp = os.path.join(self.dp_out, fn) return fp def stack_forward(self, prompt2, seed2): # Save preview images, prompts and seeds into dictionary for stacking self.dict_multi_trans[self.current_timestamp] = generate_list_output(self.prompt1, self.prompt2, self.seed1, self.seed2, self.list_fp_imgs_current) self.dict_multi_trans_include[self.current_timestamp] = True self.lb.swap_forward() list_out = [self.list_fp_imgs_current[-1]] list_out.extend([self.fp_img_empty]*4) list_out.append(prompt2) list_out.append(seed2) list_out.append("") list_out.append(np.random.randint(0, 10000000)) list_out_multi_tab = self.update_trans_stacks() list_out.extend(list_out_multi_tab) # self.nmb_trans_stack = len(self.dict_multi_trans_include) return list_out def stack_movie(self): # collect all that are in... list_fp_movies = [] for timestamp in self.multi_trans_currently_shown: if timestamp is not None: list_fp_movies.append(self.get_fp_movie(timestamp)) fp_stacked = self.get_fp_movie(get_time('second'), True) concatenate_movies(fp_stacked, list_fp_movies) return fp_stacked def get_state_dict(self): state_dict = {} grab_vars = ['prompt1', 'prompt2', 'seed1', 'seed2', 'height', 'width', 'num_inference_steps', 'depth_strength', 'guidance_scale', 'guidance_scale_mid_damper', 'mid_compression_scaler'] for v in grab_vars: state_dict[v] = getattr(self, v) return state_dict def update_trans_stacks(self): print("Updating transition stack...") self.multi_trans_currently_shown = [] list_output = [] # Figure out which transitions should be shown for timestamp in self.dict_multi_trans_include.keys(): if len(self.multi_trans_currently_shown) >= self.nmb_trans_stack: continue if self.dict_multi_trans_include[timestamp]: last_timestamp_vals = self.dict_multi_trans[timestamp] list_output.extend(self.dict_multi_trans[timestamp]) self.multi_trans_currently_shown.append(timestamp) print(f"including timestamp: {timestamp}") # Fill with empty images if below nmb_trans_stack nmb_empty_missing = self.nmb_trans_stack - len(self.multi_trans_currently_shown) for i in range(nmb_empty_missing): list_output.extend([gr.update(visible=False)]*len(last_timestamp_vals)) self.multi_trans_currently_shown.append(None) return list_output def remove_trans(self, idx_row): idx_row = int(idx_row) # do removal... if idx_row < len(self.multi_trans_currently_shown): timestamp = self.multi_trans_currently_shown[idx_row] if timestamp in self.dict_multi_trans_include.keys(): self.dict_multi_trans_include[timestamp] = False print(f"remove_trans called: {timestamp}") else: print(f"remove_trans called: idx_row too large {idx_row}") return self.update_trans_stacks() def get_img_rand(): return (255*np.random.rand(self.height,self.width,3)).astype(np.uint8) def generate_list_output( prompt1, prompt2, seed1, seed2, list_fp_imgs, ): list_output = [] list_output.append(prompt1) list_output.append(prompt2) list_output.append(seed1) list_output.append(seed2) for fp_img in list_fp_imgs: list_output.append(fp_img) return list_output if __name__ == "__main__": # fp_ckpt = "../stable_diffusion_models/ckpt/v2-1_768-ema-pruned.ckpt" fp_ckpt = "../stable_diffusion_models/ckpt/v2-1_512-ema-pruned.ckpt" sdh = StableDiffusionHolder(fp_ckpt) self = BlendingFrontend(sdh) # Yes this is possible in python and yes it is an awesome trick # self = BlendingFrontend(None) # Yes this is possible in python and yes it is an awesome trick dict_ui_elem = {} with gr.Blocks() as demo: with gr.Row(): prompt1 = gr.Textbox(label="prompt 1") prompt2 = gr.Textbox(label="prompt 2") with gr.Row(): duration_compute = gr.Slider(5, 45, self.t_compute_max_allowed, step=1, label='compute budget for transition (seconds)', interactive=True) duration_video = gr.Slider(0.1, 30, self.duration_video, step=0.1, label='result video duration (seconds)', interactive=True) height = gr.Slider(256, 2048, self.height, step=128, label='height', interactive=True) width = gr.Slider(256, 2048, self.width, step=128, label='width', interactive=True) with gr.Accordion("Advanced Settings (click to expand)", open=False): with gr.Accordion("Diffusion settings", open=True): with gr.Row(): num_inference_steps = gr.Slider(5, 100, self.num_inference_steps, step=1, label='num_inference_steps', interactive=True) guidance_scale = gr.Slider(1, 25, self.guidance_scale, step=0.1, label='guidance_scale', interactive=True) negative_prompt = gr.Textbox(label="negative prompt") with gr.Accordion("Seeds control", open=True): with gr.Row(): seed1 = gr.Number(420, label="seed 1", interactive=True) b_newseed1 = gr.Button("randomize seed 1", variant='secondary') seed2 = gr.Number(420, label="seed 2", interactive=True) b_newseed2 = gr.Button("randomize seed 2", variant='secondary') with gr.Accordion("Crossfeeding for last image", open=True): with gr.Row(): branch1_influence = gr.Slider(0.0, 1.0, self.branch1_influence, step=0.01, label='crossfeed power', interactive=True) branch1_max_depth_influence = gr.Slider(0.0, 1.0, self.branch1_max_depth_influence, step=0.01, label='crossfeed range', interactive=True) branch1_influence_decay = gr.Slider(0.0, 1.0, self.branch1_influence_decay, step=0.01, label='crossfeed decay', interactive=True) with gr.Accordion("Transition settings", open=True): with gr.Row(): depth_strength = gr.Slider(0.01, 0.99, self.depth_strength, step=0.01, label='depth_strength', interactive=True) guidance_scale_mid_damper = gr.Slider(0.01, 2.0, self.guidance_scale_mid_damper, step=0.01, label='guidance_scale_mid_damper', interactive=True) parental_influence = gr.Slider(0.0, 1.0, self.parental_influence, step=0.01, label='parental power', interactive=True) parental_max_depth_influence = gr.Slider(0.0, 1.0, self.parental_max_depth_influence, step=0.01, label='parental range', interactive=True) parental_influence_decay = gr.Slider(0.0, 1.0, self.parental_influence_decay, step=0.01, label='parental decay', interactive=True) with gr.Row(): b_compute1 = gr.Button('compute first image', variant='primary') b_compute_transition = gr.Button('compute transition', variant='primary') b_compute2 = gr.Button('compute last image', variant='primary') with gr.Row(): img1 = gr.Image(label="1/5") img2 = gr.Image(label="2/5") img3 = gr.Image(label="3/5") img4 = gr.Image(label="4/5") img5 = gr.Image(label="5/5") with gr.Row(): vid_transition = gr.Video() # Collect all UI elemts in list to easily pass as inputs dict_ui_elem["prompt1"] = prompt1 dict_ui_elem["negative_prompt"] = negative_prompt dict_ui_elem["prompt2"] = prompt2 dict_ui_elem["duration_compute"] = duration_compute dict_ui_elem["duration_video"] = duration_video dict_ui_elem["height"] = height dict_ui_elem["width"] = width dict_ui_elem["depth_strength"] = depth_strength dict_ui_elem["branch1_influence"] = branch1_influence dict_ui_elem["branch1_max_depth_influence"] = branch1_max_depth_influence dict_ui_elem["branch1_influence_decay"] = branch1_influence_decay dict_ui_elem["num_inference_steps"] = num_inference_steps dict_ui_elem["guidance_scale"] = guidance_scale dict_ui_elem["guidance_scale_mid_damper"] = guidance_scale_mid_damper dict_ui_elem["seed1"] = seed1 dict_ui_elem["seed2"] = seed2 dict_ui_elem["parental_max_depth_influence"] = parental_max_depth_influence dict_ui_elem["parental_influence"] = parental_influence dict_ui_elem["parental_influence_decay"] = parental_influence_decay # Convert to list, as gradio doesn't seem to accept dicts list_ui_elem = [] list_ui_keys = [] for k in dict_ui_elem.keys(): list_ui_elem.append(dict_ui_elem[k]) list_ui_keys.append(k) self.list_ui_keys = list_ui_keys b_newseed1.click(self.randomize_seed1, outputs=seed1) b_newseed2.click(self.randomize_seed2, outputs=seed2) b_compute1.click(self.compute_img1, inputs=list_ui_elem, outputs=[img1, img2, img3, img4, img5]) b_compute2.click(self.compute_img2, inputs=list_ui_elem, outputs=[img2, img3, img4, img5]) b_compute_transition.click(self.compute_transition, inputs=list_ui_elem, outputs=[img2, img3, img4, vid_transition]) demo.launch(share=self.share, inbrowser=True, inline=False)