From 3de20215426882ad769971caf9ba963270ab70bb Mon Sep 17 00:00:00 2001 From: DGX Date: Wed, 21 Feb 2024 15:22:34 +0000 Subject: [PATCH] simple gradio interface for saving jsons --- .gitignore | 1 + examples/multi_trans_json.py | 75 +++++ latentblending/gradio_ui.py | 543 +++++++---------------------------- 3 files changed, 174 insertions(+), 445 deletions(-) create mode 100644 examples/multi_trans_json.py diff --git a/.gitignore b/.gitignore index 5e629be..de29b3f 100644 --- a/.gitignore +++ b/.gitignore @@ -7,6 +7,7 @@ __pycache__/ *.so # Distribution / packaging +*.json .Python build/ develop-eggs/ diff --git a/examples/multi_trans_json.py b/examples/multi_trans_json.py new file mode 100644 index 0000000..fafaffb --- /dev/null +++ b/examples/multi_trans_json.py @@ -0,0 +1,75 @@ +import torch +import warnings +from diffusers import AutoPipelineForText2Image +from latentblending.blending_engine import BlendingEngine +from lunar_tools import concatenate_movies +import numpy as np +torch.set_grad_enabled(False) +torch.backends.cudnn.benchmark = False +warnings.filterwarnings('ignore') + +import json +# %% First let us spawn a stable diffusion holder. Uncomment your version of choice. +# pretrained_model_name_or_path = "stabilityai/stable-diffusion-xl-base-1.0" +pretrained_model_name_or_path = "stabilityai/sdxl-turbo" + +pipe = AutoPipelineForText2Image.from_pretrained(pretrained_model_name_or_path, torch_dtype=torch.float16, variant="fp16") +pipe.to('cuda') +be = BlendingEngine(pipe, do_compile=False) + +fp_movie = f'test.mp4' +fp_json = "movie_240221_1520.json" +duration_single_trans = 10 + +# Load the JSON data from the file +with open(fp_json, 'r') as file: + data = json.load(file) + +# Set up width, height, num_inference steps +width = data[0]["width"] +height = data[0]["height"] +num_inference_steps = data[0]["num_inference_steps"] + +be.set_dimensions((width, height)) +be.set_num_inference_steps(num_inference_steps) + +# Initialize lists for prompts, negative prompts, and seeds +list_prompts = [] +list_negative_prompts = [] +list_seeds = [] + +# Extract prompts, negative prompts, and seeds from the data +for item in data[1:]: # Skip the first item as it contains settings + list_prompts.append(item["prompt"]) + list_negative_prompts.append(item["negative_prompt"]) + list_seeds.append(item["seed"]) + + +list_movie_parts = [] +for i in range(len(list_prompts) - 1): + # For a multi transition we can save some computation time and recycle the latents + if i == 0: + be.set_prompt1(list_prompts[i]) + be.set_negative_prompt(list_negative_prompts[i]) + be.set_prompt2(list_prompts[i + 1]) + recycle_img1 = False + else: + be.swap_forward() + be.set_negative_prompt(list_negative_prompts[i+1]) + be.set_prompt2(list_prompts[i + 1]) + recycle_img1 = True + + fp_movie_part = f"tmp_part_{str(i).zfill(3)}.mp4" + fixed_seeds = list_seeds[i:i + 2] + # Run latent blending + be.run_transition( + recycle_img1=recycle_img1, + fixed_seeds=fixed_seeds) + + # Save movie + be.write_movie_transition(fp_movie_part, duration_single_trans) + list_movie_parts.append(fp_movie_part) + +# Finally, concatente the result +concatenate_movies(fp_movie, list_movie_parts) +print(f"DONE! MOVIE SAVED IN {fp_movie}") \ No newline at end of file diff --git a/latentblending/gradio_ui.py b/latentblending/gradio_ui.py index 106cf54..2b6d491 100644 --- a/latentblending/gradio_ui.py +++ b/latentblending/gradio_ui.py @@ -1,18 +1,3 @@ -# 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 import torch torch.backends.cudnn.benchmark = False @@ -20,481 +5,149 @@ torch.set_grad_enabled(False) import numpy as np import warnings warnings.filterwarnings('ignore') -import warnings from tqdm.auto import tqdm from PIL import Image -from movie_util import MovieSaver, concatenate_movies -from latent_blending import LatentBlending -from stable_diffusion_holder import StableDiffusionHolder import gradio as gr -from dotenv import find_dotenv, load_dotenv import shutil import uuid -from utils import get_time, add_frames_linear_interp -from huggingface_hub import hf_hub_download +from diffusers import AutoPipelineForText2Image +from latentblending.blending_engine import BlendingEngine +import datetime + +warnings.filterwarnings('ignore') +torch.set_grad_enabled(False) +torch.backends.cudnn.benchmark = False +import json + class BlendingFrontend(): def __init__( self, - sdh, + be, share=False): r""" Gradio Helper Class to collect UI data and start latent blending. Args: - sdh: - StableDiffusionHolder + be: + Blendingengine share: bool Set true to get a shareable gradio link (e.g. for running a remote server) """ + self.be = be self.share = share # UI Defaults - self.num_inference_steps = 30 - self.depth_strength = 0.25 self.seed1 = 420 self.seed2 = 420 self.prompt1 = "" self.prompt2 = "" self.negative_prompt = "" - self.fps = 30 - self.duration_video = 8 - self.t_compute_max_allowed = 10 - - self.lb = LatentBlending(sdh) - self.lb.sdh.num_inference_steps = self.num_inference_steps - self.init_parameters_from_lb() - self.init_save_dir() # Vars - self.list_fp_imgs_current = [] - self.recycle_img1 = False - self.recycle_img2 = False - self.list_all_segments = [] - self.dp_session = "" - self.user_id = None + self.prompt = None + self.negative_prompt = None + self.list_seeds = [] + self.idx_movie = 0 + self.data = [] - def init_parameters_from_lb(self): - r""" - Automatically init parameters from latentblending instance - """ - self.height = self.lb.sdh.height - self.width = self.lb.sdh.width - self.guidance_scale = self.lb.guidance_scale - self.guidance_scale_mid_damper = self.lb.guidance_scale_mid_damper - self.mid_compression_scaler = self.lb.mid_compression_scaler - self.branch1_crossfeed_power = self.lb.branch1_crossfeed_power - self.branch1_crossfeed_range = self.lb.branch1_crossfeed_range - self.branch1_crossfeed_decay = self.lb.branch1_crossfeed_decay - self.parental_crossfeed_power = self.lb.parental_crossfeed_power - self.parental_crossfeed_range = self.lb.parental_crossfeed_range - self.parental_crossfeed_power_decay = self.lb.parental_crossfeed_power_decay + def take_image0(self): + return self.take_image(0) + + def take_image1(self): + return self.take_image(1) + + def take_image2(self): + return self.take_image(2) + + def take_image3(self): + return self.take_image(3) + - def init_save_dir(self): - r""" - Initializes the directory where stuff is being saved. - You can specify this directory in a ".env" file in your latentblending root, setting - DIR_OUT='/path/to/saving' - """ - load_dotenv(find_dotenv(), verbose=False) - self.dp_out = os.getenv("DIR_OUT") - if self.dp_out is None: - self.dp_out = "" - self.dp_imgs = os.path.join(self.dp_out, "imgs") - os.makedirs(self.dp_imgs, exist_ok=True) - self.dp_movies = os.path.join(self.dp_out, "movies") - os.makedirs(self.dp_movies, exist_ok=True) - self.save_empty_image() + def take_image(self, id_img): + if self.prompt is None: + print("Cannot take because no prompt was set!") + return [None, None, None, None, ""] + if self.idx_movie == 0: + current_time = datetime.datetime.now() + self.fp_out = "movie_" + current_time.strftime("%y%m%d_%H%M") + ".json" + self.data.append({"settings": "sdxl", "width": bf.be.dh.width_img, "height": self.be.dh.height_img, "num_inference_steps": self.be.dh.num_inference_steps}) - def save_empty_image(self): - r""" - Saves an empty/black dummy image. - """ - self.fp_img_empty = os.path.join(self.dp_imgs, 'empty.jpg') - Image.fromarray(np.zeros((self.height, self.width, 3), dtype=np.uint8)).save(self.fp_img_empty, quality=5) + seed = self.list_seeds[id_img] + + self.data.append({"iteration": self.idx_movie, "seed": seed, "prompt": self.prompt, "negative_prompt": self.negative_prompt}) - def randomize_seed1(self): - r""" - Randomizes the first seed - """ - seed = np.random.randint(0, 10000000) - self.seed1 = int(seed) - print(f"randomize_seed1: new seed = {self.seed1}") - return seed + # Write the data list to a JSON file + with open(self.fp_out, 'w') as f: + json.dump(self.data, f, indent=4) - def randomize_seed2(self): - r""" - Randomizes the second seed - """ - seed = np.random.randint(0, 10000000) - self.seed2 = int(seed) - print(f"randomize_seed2: new seed = {self.seed2}") - return seed + self.idx_movie += 1 + self.prompt = None + return [None, None, None, None, ""] - def setup_lb(self, list_ui_vals): - r""" - Sets all parameters from the UI. Since gradio does not support to pass dictionaries, - we have to instead pass keys (list_ui_keys, global) and values (list_ui_vals) - """ - # Collect latent blending variables - self.lb.set_width(list_ui_vals[list_ui_keys.index('width')]) - self.lb.set_height(list_ui_vals[list_ui_keys.index('height')]) - self.lb.set_prompt1(list_ui_vals[list_ui_keys.index('prompt1')]) - self.lb.set_prompt2(list_ui_vals[list_ui_keys.index('prompt2')]) - self.lb.set_negative_prompt(list_ui_vals[list_ui_keys.index('negative_prompt')]) - self.lb.guidance_scale = list_ui_vals[list_ui_keys.index('guidance_scale')] - self.lb.guidance_scale_mid_damper = list_ui_vals[list_ui_keys.index('guidance_scale_mid_damper')] - self.t_compute_max_allowed = list_ui_vals[list_ui_keys.index('duration_compute')] - self.lb.num_inference_steps = list_ui_vals[list_ui_keys.index('num_inference_steps')] - self.lb.sdh.num_inference_steps = list_ui_vals[list_ui_keys.index('num_inference_steps')] - self.duration_video = list_ui_vals[list_ui_keys.index('duration_video')] - self.lb.seed1 = list_ui_vals[list_ui_keys.index('seed1')] - self.lb.seed2 = list_ui_vals[list_ui_keys.index('seed2')] - self.lb.branch1_crossfeed_power = list_ui_vals[list_ui_keys.index('branch1_crossfeed_power')] - self.lb.branch1_crossfeed_range = list_ui_vals[list_ui_keys.index('branch1_crossfeed_range')] - self.lb.branch1_crossfeed_decay = list_ui_vals[list_ui_keys.index('branch1_crossfeed_decay')] - self.lb.parental_crossfeed_power = list_ui_vals[list_ui_keys.index('parental_crossfeed_power')] - self.lb.parental_crossfeed_range = list_ui_vals[list_ui_keys.index('parental_crossfeed_range')] - self.lb.parental_crossfeed_power_decay = list_ui_vals[list_ui_keys.index('parental_crossfeed_power_decay')] - self.num_inference_steps = list_ui_vals[list_ui_keys.index('num_inference_steps')] - self.depth_strength = list_ui_vals[list_ui_keys.index('depth_strength')] - if len(list_ui_vals[list_ui_keys.index('user_id')]) > 1: - self.user_id = list_ui_vals[list_ui_keys.index('user_id')] - else: - # generate new user id - self.user_id = uuid.uuid4().hex - print(f"made new user_id: {self.user_id} at {get_time('second')}") - - def save_latents(self, fp_latents, list_latents): - r""" - Saves a latent trajectory on disk, in npy format. - """ - list_latents_cpu = [l.cpu().numpy() for l in list_latents] - np.save(fp_latents, list_latents_cpu) - - def load_latents(self, fp_latents): - r""" - Loads a latent trajectory from disk, converts to torch tensor. - """ - list_latents_cpu = np.load(fp_latents) - list_latents = [torch.from_numpy(l).to(self.lb.device) for l in list_latents_cpu] - return list_latents - - def compute_img1(self, *args): - r""" - Computes the first transition image and returns it for display. - Sets all other transition images and last image to empty (as they are obsolete with this operation) - """ - list_ui_vals = args - self.setup_lb(list_ui_vals) - fp_img1 = os.path.join(self.dp_imgs, f"img1_{self.user_id}") - img1 = Image.fromarray(self.lb.compute_latents1(return_image=True)) - img1.save(fp_img1 + ".jpg") - self.save_latents(fp_img1 + ".npy", self.lb.tree_latents[0]) - self.recycle_img1 = True - self.recycle_img2 = False - return [fp_img1 + ".jpg", self.fp_img_empty, self.fp_img_empty, self.fp_img_empty, self.fp_img_empty, self.user_id] - - def compute_img2(self, *args): - r""" - Computes the last transition image and returns it for display. - Sets all other transition images to empty (as they are obsolete with this operation) - """ - if not os.path.isfile(os.path.join(self.dp_imgs, f"img1_{self.user_id}.jpg")): # don't do anything - return [self.fp_img_empty, self.fp_img_empty, self.fp_img_empty, self.fp_img_empty, self.user_id] - list_ui_vals = args - self.setup_lb(list_ui_vals) - - self.lb.tree_latents[0] = self.load_latents(os.path.join(self.dp_imgs, f"img1_{self.user_id}.npy")) - fp_img2 = os.path.join(self.dp_imgs, f"img2_{self.user_id}") - img2 = Image.fromarray(self.lb.compute_latents2(return_image=True)) - img2.save(fp_img2 + '.jpg') - self.save_latents(fp_img2 + ".npy", self.lb.tree_latents[-1]) - self.recycle_img2 = True - # fixme save seeds. change filenames? - return [self.fp_img_empty, self.fp_img_empty, self.fp_img_empty, fp_img2 + ".jpg", self.user_id] - - def compute_transition(self, *args): - r""" - Computes transition images and movie. - """ - list_ui_vals = args - self.setup_lb(list_ui_vals) - print("STARTING TRANSITION...") - fixed_seeds = [self.seed1, self.seed2] - # Inject loaded latents (other user interference) - self.lb.tree_latents[0] = self.load_latents(os.path.join(self.dp_imgs, f"img1_{self.user_id}.npy")) - self.lb.tree_latents[-1] = self.load_latents(os.path.join(self.dp_imgs, f"img2_{self.user_id}.npy")) - imgs_transition = self.lb.run_transition( - recycle_img1=self.recycle_img1, - recycle_img2=self.recycle_img2, - num_inference_steps=self.num_inference_steps, - depth_strength=self.depth_strength, - t_compute_max_allowed=self.t_compute_max_allowed, - fixed_seeds=fixed_seeds) - print(f"Latent Blending pass finished ({get_time('second')}). 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 - current_timestamp = get_time('second') - self.list_fp_imgs_current = [] - for i in range(len(list_imgs_preview)): - fp_img = os.path.join(self.dp_imgs, f"img_preview_{i}_{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 - self.fp_movie = self.get_fp_video_last() - if os.path.isfile(self.fp_movie): - os.remove(self.fp_movie) - ms = MovieSaver(self.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 + [self.fp_movie] - return list_return - - def stack_forward(self, prompt2, seed2): - r""" - Allows to generate multi-segment movies. Sets last image -> first image with all - relevant parameters. - """ - # Save preview images, prompts and seeds into dictionary for stacking - if len(self.list_all_segments) == 0: - timestamp_session = get_time('second') - self.dp_session = os.path.join(self.dp_out, f"session_{timestamp_session}") - os.makedirs(self.dp_session) - - idx_segment = len(self.list_all_segments) - dp_segment = os.path.join(self.dp_session, f"segment_{str(idx_segment).zfill(3)}") - - self.list_all_segments.append(dp_segment) - self.lb.write_imgs_transition(dp_segment) - - fp_movie_last = self.get_fp_video_last() - fp_movie_next = self.get_fp_video_next() - - shutil.copyfile(fp_movie_last, fp_movie_next) - - self.lb.tree_latents[0] = self.load_latents(os.path.join(self.dp_imgs, f"img1_{self.user_id}.npy")) - self.lb.tree_latents[-1] = self.load_latents(os.path.join(self.dp_imgs, f"img2_{self.user_id}.npy")) - self.lb.swap_forward() - - shutil.copyfile(os.path.join(self.dp_imgs, f"img2_{self.user_id}.npy"), os.path.join(self.dp_imgs, f"img1_{self.user_id}.npy")) - fp_multi = self.multi_concat() - list_out = [fp_multi] - - list_out.extend([os.path.join(self.dp_imgs, f"img2_{self.user_id}.jpg")]) - list_out.extend([self.fp_img_empty] * 4) - list_out.append(gr.update(interactive=False, value=prompt2)) - list_out.append(gr.update(interactive=False, value=seed2)) - list_out.append("") - list_out.append(np.random.randint(0, 10000000)) - print(f"stack_forward: fp_multi {fp_multi}") - return list_out - - def multi_concat(self): - r""" - Concatentates all stacked segments into one long movie. - """ - list_fp_movies = self.get_fp_video_all() - # Concatenate movies and save - fp_final = os.path.join(self.dp_session, f"concat_{self.user_id}.mp4") - concatenate_movies(fp_final, list_fp_movies) - return fp_final - - def get_fp_video_all(self): - r""" - Collects all stacked movie segments. - """ - list_all = os.listdir(self.dp_movies) - str_beg = f"movie_{self.user_id}_" - list_user = [l for l in list_all if str_beg in l] - list_user.sort() - list_user = [os.path.join(self.dp_movies, l) for l in list_user] - return list_user - - def get_fp_video_next(self): - r""" - Gets the filepath of the next movie segment. - """ - list_videos = self.get_fp_video_all() - if len(list_videos) == 0: - idx_next = 0 - else: - idx_next = len(list_videos) - fp_video_next = os.path.join(self.dp_movies, f"movie_{self.user_id}_{str(idx_next).zfill(3)}.mp4") - return fp_video_next - - def get_fp_video_last(self): - r""" - Gets the current video that was saved. - """ - fp_video_last = os.path.join(self.dp_movies, f"last_{self.user_id}.mp4") - return fp_video_last + def compute_imgs(self, prompt, negative_prompt): + self.prompt = prompt + self.negative_prompt = negative_prompt + self.be.set_prompt1(prompt) + self.be.set_prompt2(prompt) + self.be.set_negative_prompt(negative_prompt) + self.list_seeds = [] + self.list_images = [] + for i in range(4): + seed = np.random.randint(0, 1000000000) + self.be.seed1 = seed + self.list_seeds.append(seed) + img = self.be.compute_latents1(return_image=True) + self.list_images.append(img) + return self.list_images + + if __name__ == "__main__": -# fp_ckpt = hf_hub_download(repo_id="stabilityai/stable-diffusion-2-1-base", filename="v2-1_512-ema-pruned.ckpt") - fp_ckpt = hf_hub_download(repo_id="stabilityai/stable-diffusion-2-1", filename="v2-1_768-ema-pruned.ckpt") - bf = BlendingFrontend(StableDiffusionHolder(fp_ckpt)) - # self = BlendingFrontend(None) + + width = 786 + height = 1024 + num_inference_steps = 4 + + pipe = AutoPipelineForText2Image.from_pretrained("stabilityai/sdxl-turbo", torch_dtype=torch.float16, variant="fp16") + # pipe = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16, variant="fp16") + pipe.to("cuda") + + be = BlendingEngine(pipe) + be.set_dimensions((width, height)) + be.set_num_inference_steps(num_inference_steps) + + bf = BlendingFrontend(be) with gr.Blocks() as demo: - gr.HTML("""

Latent Blending

-

Create butter-smooth transitions between prompts, powered by stable diffusion

-

For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings. -
- -Duplicate Space -

""") with gr.Row(): - prompt1 = gr.Textbox(label="prompt 1") - prompt2 = gr.Textbox(label="prompt 2") + prompt = gr.Textbox(label="prompt") + negative_prompt = gr.Textbox(label="negative prompt") with gr.Row(): - duration_compute = gr.Slider(10, 25, bf.t_compute_max_allowed, step=1, label='waiting time', interactive=True) - duration_video = gr.Slider(1, 100, bf.duration_video, step=0.1, label='video duration', interactive=True) - height = gr.Slider(256, 1024, bf.height, step=128, label='height', interactive=True) - width = gr.Slider(256, 1024, bf.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, bf.num_inference_steps, step=1, label='num_inference_steps', interactive=True) - guidance_scale = gr.Slider(1, 25, bf.guidance_scale, step=0.1, label='guidance_scale', interactive=True) - negative_prompt = gr.Textbox(label="negative prompt") - - with gr.Accordion("Seed control: adjust seeds for first and last images", open=True): - with gr.Row(): - b_newseed1 = gr.Button("randomize seed 1", variant='secondary') - seed1 = gr.Number(bf.seed1, label="seed 1", interactive=True) - seed2 = gr.Number(bf.seed2, label="seed 2", interactive=True) - b_newseed2 = gr.Button("randomize seed 2", variant='secondary') - - with gr.Accordion("Last image crossfeeding.", open=True): - with gr.Row(): - branch1_crossfeed_power = gr.Slider(0.0, 1.0, bf.branch1_crossfeed_power, step=0.01, label='branch1 crossfeed power', interactive=True) - branch1_crossfeed_range = gr.Slider(0.0, 1.0, bf.branch1_crossfeed_range, step=0.01, label='branch1 crossfeed range', interactive=True) - branch1_crossfeed_decay = gr.Slider(0.0, 1.0, bf.branch1_crossfeed_decay, step=0.01, label='branch1 crossfeed decay', interactive=True) - - with gr.Accordion("Transition settings", open=True): - with gr.Row(): - parental_crossfeed_power = gr.Slider(0.0, 1.0, bf.parental_crossfeed_power, step=0.01, label='parental crossfeed power', interactive=True) - parental_crossfeed_range = gr.Slider(0.0, 1.0, bf.parental_crossfeed_range, step=0.01, label='parental crossfeed range', interactive=True) - parental_crossfeed_power_decay = gr.Slider(0.0, 1.0, bf.parental_crossfeed_power_decay, step=0.01, label='parental crossfeed decay', interactive=True) - with gr.Row(): - depth_strength = gr.Slider(0.01, 0.99, bf.depth_strength, step=0.01, label='depth_strength', interactive=True) - guidance_scale_mid_damper = gr.Slider(0.01, 2.0, bf.guidance_scale_mid_damper, step=0.01, label='guidance_scale_mid_damper', interactive=True) + b_compute = gr.Button('compute new images', variant='primary') with gr.Row(): - b_compute1 = gr.Button('step1: compute first image', variant='primary') - b_compute2 = gr.Button('step2: compute last image', variant='primary') - b_compute_transition = gr.Button('step3: compute transition', variant='primary') + with gr.Column(): + img0 = gr.Image(label="seed1") + b_take0 = gr.Button('take', variant='primary') + with gr.Column(): + img1 = gr.Image(label="seed2") + b_take1 = gr.Button('take', variant='primary') + with gr.Column(): + img2 = gr.Image(label="seed3") + b_take2 = gr.Button('take', variant='primary') + with gr.Column(): + img3 = gr.Image(label="seed4") + b_take3 = gr.Button('take', variant='primary') - with gr.Row(): - img1 = gr.Image(label="1/5") - img2 = gr.Image(label="2/5", show_progress=False) - img3 = gr.Image(label="3/5", show_progress=False) - img4 = gr.Image(label="4/5", show_progress=False) - img5 = gr.Image(label="5/5") - - with gr.Row(): - vid_single = gr.Video(label="current single trans") - vid_multi = gr.Video(label="concatented multi trans") - - with gr.Row(): - b_stackforward = gr.Button('append last movie segment (left) to multi movie (right)', variant='primary') - - with gr.Row(): - gr.Markdown( - """ - # Parameters - ## Main - - waiting time: set your waiting time for the transition. high values = better quality - - video duration: seconds per segment - - height/width: in pixels - - ## Diffusion settings - - num_inference_steps: number of diffusion steps - - guidance_scale: latent blending seems to prefer lower values here - - negative prompt: enter negative prompt here, applied for all images - - ## Last image crossfeeding - - branch1_crossfeed_power: Controls the level of cross-feeding between the first and last image branch. For preserving structures. - - branch1_crossfeed_range: Sets the duration of active crossfeed during development. High values enforce strong structural similarity. - - branch1_crossfeed_decay: Sets decay for branch1_crossfeed_power. Lower values make the decay stronger across the range. - - ## Transition settings - - parental_crossfeed_power: Similar to branch1_crossfeed_power, however applied for the images withinin the transition. - - parental_crossfeed_range: Similar to branch1_crossfeed_range, however applied for the images withinin the transition. - - parental_crossfeed_power_decay: Similar to branch1_crossfeed_decay, however applied for the images withinin the transition. - - depth_strength: Determines when the blending process will begin in terms of diffusion steps. Low values more inventive but can cause motion. - - guidance_scale_mid_damper: Decreases the guidance scale in the middle of a transition. - """) - - with gr.Row(): - user_id = gr.Textbox(label="user id", interactive=False) - - # Collect all UI elemts in list to easily pass as inputs in gradio - dict_ui_elem = {} - 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_crossfeed_power"] = branch1_crossfeed_power - dict_ui_elem["branch1_crossfeed_range"] = branch1_crossfeed_range - dict_ui_elem["branch1_crossfeed_decay"] = branch1_crossfeed_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_crossfeed_range"] = parental_crossfeed_range - dict_ui_elem["parental_crossfeed_power"] = parental_crossfeed_power - dict_ui_elem["parental_crossfeed_power_decay"] = parental_crossfeed_power_decay - dict_ui_elem["user_id"] = user_id - - # Convert to list, as gradio doesn't seem to accept dicts - list_ui_vals = [] - list_ui_keys = [] - for k in dict_ui_elem.keys(): - list_ui_vals.append(dict_ui_elem[k]) - list_ui_keys.append(k) - bf.list_ui_keys = list_ui_keys - - b_newseed1.click(bf.randomize_seed1, outputs=seed1) - b_newseed2.click(bf.randomize_seed2, outputs=seed2) - b_compute1.click(bf.compute_img1, inputs=list_ui_vals, outputs=[img1, img2, img3, img4, img5, user_id]) - b_compute2.click(bf.compute_img2, inputs=list_ui_vals, outputs=[img2, img3, img4, img5, user_id]) - b_compute_transition.click(bf.compute_transition, - inputs=list_ui_vals, - outputs=[img2, img3, img4, vid_single]) - - b_stackforward.click(bf.stack_forward, - inputs=[prompt2, seed2], - outputs=[vid_multi, img1, img2, img3, img4, img5, prompt1, seed1, prompt2]) + b_compute.click(bf.compute_imgs, inputs=[prompt, negative_prompt], outputs=[img0, img1, img2, img3]) + b_take0.click(bf.take_image0, outputs=[img0, img1, img2, img3, prompt]) + b_take1.click(bf.take_image1, outputs=[img0, img1, img2, img3, prompt]) + b_take2.click(bf.take_image2, outputs=[img0, img1, img2, img3, prompt]) + b_take3.click(bf.take_image3, outputs=[img0, img1, img2, img3, prompt]) demo.launch(share=bf.share, inbrowser=True, inline=False)