latentblending/gradio_ui.py

501 lines
23 KiB
Python
Raw Normal View History

2023-01-08 09:33:45 +00:00
# Copyright 2022 Lunar Ring. All rights reserved.
2023-01-11 11:58:59 +00:00
# Written by Johannes Stelzer, email stelzer@lunar-ring.ai twitter @j_stelzer
2023-01-08 09:33:45 +00:00
#
# 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.
2023-02-22 09:15:03 +00:00
import os
2023-01-08 09:33:45 +00:00
import torch
torch.backends.cudnn.benchmark = False
2023-02-22 09:15:03 +00:00
torch.set_grad_enabled(False)
2023-01-08 09:33:45 +00:00
import numpy as np
import warnings
warnings.filterwarnings('ignore')
import warnings
from tqdm.auto import tqdm
from PIL import Image
2023-01-15 15:52:42 +00:00
from movie_util import MovieSaver, concatenate_movies
2023-02-22 09:15:03 +00:00
from latent_blending import LatentBlending
2023-01-08 09:33:45 +00:00
from stable_diffusion_holder import StableDiffusionHolder
import gradio as gr
2023-02-16 10:48:45 +00:00
from dotenv import find_dotenv, load_dotenv
2023-02-18 06:56:30 +00:00
import shutil
import random
2023-02-22 09:15:03 +00:00
from utils import get_time, add_frames_linear_interp
from huggingface_hub import hf_hub_download
2023-01-15 11:02:11 +00:00
2023-01-08 09:33:45 +00:00
class BlendingFrontend():
2023-02-22 09:15:03 +00:00
def __init__(
self,
sdh,
share=False):
r"""
Gradio Helper Class to collect UI data and start latent blending.
Args:
sdh:
StableDiffusionHolder
share: bool
Set true to get a shareable gradio link (e.g. for running a remote server)
"""
self.share = share
# UI Defaults
2023-02-18 06:56:30 +00:00
self.num_inference_steps = 30
2023-01-08 09:33:45 +00:00
self.depth_strength = 0.25
2023-02-18 06:56:30 +00:00
self.seed1 = 420
2023-01-08 09:33:45 +00:00
self.seed2 = 420
2023-01-08 10:48:44 +00:00
self.prompt1 = ""
self.prompt2 = ""
self.negative_prompt = ""
2023-01-09 08:58:26 +00:00
self.fps = 30
2023-02-22 09:15:03 +00:00
self.duration_video = 8
2023-02-16 10:48:45 +00:00
self.t_compute_max_allowed = 10
2023-02-22 09:15:03 +00:00
self.lb = LatentBlending(sdh)
self.lb.sdh.num_inference_steps = self.num_inference_steps
self.init_parameters_from_lb()
self.init_save_dir()
# Vars
2023-01-15 15:52:42 +00:00
self.list_fp_imgs_current = []
2023-02-17 09:50:57 +00:00
self.recycle_img1 = False
self.recycle_img2 = False
2023-02-20 07:29:21 +00:00
self.list_all_segments = []
self.dp_session = ""
self.user_id = None
2023-02-22 09:15:03 +00:00
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
2023-02-16 10:48:45 +00:00
def init_save_dir(self):
2023-02-22 09:15:03 +00:00
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)
2023-02-20 07:29:21 +00:00
self.dp_out = os.getenv("DIR_OUT")
2023-02-19 14:46:02 +00:00
if self.dp_out is None:
2023-02-16 10:48:45 +00:00
self.dp_out = ""
2023-02-18 06:56:30 +00:00
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)
2023-02-22 09:15:03 +00:00
self.save_empty_image()
2023-02-15 17:21:00 +00:00
def save_empty_image(self):
2023-02-22 09:15:03 +00:00
r"""
Saves an empty/black dummy image.
"""
2023-02-18 06:56:30 +00:00
self.fp_img_empty = os.path.join(self.dp_imgs, 'empty.jpg')
2023-01-15 15:52:42 +00:00
Image.fromarray(np.zeros((self.height, self.width, 3), dtype=np.uint8)).save(self.fp_img_empty, quality=5)
2023-02-22 09:15:03 +00:00
2023-01-08 09:33:45 +00:00
def randomize_seed1(self):
2023-02-22 09:15:03 +00:00
r"""
Randomizes the first seed
"""
seed = np.random.randint(0, 10000000)
2023-02-15 17:21:00 +00:00
self.seed1 = int(seed)
2023-01-08 09:33:45 +00:00
print(f"randomize_seed1: new seed = {self.seed1}")
return seed
2023-02-22 09:15:03 +00:00
2023-01-08 09:33:45 +00:00
def randomize_seed2(self):
2023-02-22 09:15:03 +00:00
r"""
Randomizes the second seed
"""
2023-01-08 09:33:45 +00:00
seed = np.random.randint(0, 10000000)
2023-02-15 17:21:00 +00:00
self.seed2 = int(seed)
2023-01-08 09:33:45 +00:00
print(f"randomize_seed2: new seed = {self.seed2}")
return seed
2023-02-22 09:15:03 +00:00
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)
"""
2023-02-15 17:21:00 +00:00
# Collect latent blending variables
2023-02-22 09:15:03 +00:00
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 = ''.join((random.choice('ABCDEFGHIJKLMNOPQRSTUVWXYZ') for i in range(8)))
2023-02-22 09:15:03 +00:00
print(f"made new user_id: {self.user_id} at {get_time('second')}")
def save_latents(self, fp_latents, list_latents):
2023-02-22 09:15:03 +00:00
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)
2023-02-22 09:15:03 +00:00
def load_latents(self, fp_latents):
2023-02-22 09:15:03 +00:00
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
2023-02-22 09:15:03 +00:00
2023-02-15 17:21:00 +00:00
def compute_img1(self, *args):
2023-02-22 09:15:03 +00:00
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}")
2023-02-15 17:21:00 +00:00
img1 = Image.fromarray(self.lb.compute_latents1(return_image=True))
2023-02-22 09:15:03 +00:00
img1.save(fp_img1 + ".jpg")
self.save_latents(fp_img1 + ".npy", self.lb.tree_latents[0])
2023-02-17 09:50:57 +00:00
self.recycle_img1 = True
self.recycle_img2 = False
2023-02-22 09:15:03 +00:00
return [fp_img1 + ".jpg", self.fp_img_empty, self.fp_img_empty, self.fp_img_empty, self.fp_img_empty, self.user_id]
2023-02-15 17:21:00 +00:00
def compute_img2(self, *args):
2023-02-22 09:15:03 +00:00
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]
2023-02-22 09:15:03 +00:00
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}")
2023-02-15 17:21:00 +00:00
img2 = Image.fromarray(self.lb.compute_latents2(return_image=True))
2023-02-22 09:15:03 +00:00
img2.save(fp_img2 + '.jpg')
self.save_latents(fp_img2 + ".npy", self.lb.tree_latents[-1])
2023-02-17 09:50:57 +00:00
self.recycle_img2 = True
# fixme save seeds. change filenames?
2023-02-22 09:15:03 +00:00
return [self.fp_img_empty, self.fp_img_empty, self.fp_img_empty, fp_img2 + ".jpg", self.user_id]
2023-02-15 17:21:00 +00:00
def compute_transition(self, *args):
2023-02-22 09:15:03 +00:00
r"""
Computes transition images and movie.
"""
list_ui_vals = args
self.setup_lb(list_ui_vals)
2023-02-20 07:29:21 +00:00
print("STARTING TRANSITION...")
2023-01-08 09:33:45 +00:00
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"))
2023-02-15 17:21:00 +00:00
imgs_transition = self.lb.run_transition(
2023-02-22 09:15:03 +00:00
recycle_img1=self.recycle_img1,
recycle_img2=self.recycle_img2,
num_inference_steps=self.num_inference_steps,
depth_strength=self.depth_strength,
2023-02-17 09:50:57 +00:00
t_compute_max_allowed=self.t_compute_max_allowed,
2023-02-22 09:15:03 +00:00
fixed_seeds=fixed_seeds)
print(f"Latent Blending pass finished ({get_time('second')}). Resulted in {len(imgs_transition)} images")
2023-02-15 17:21:00 +00:00
# Subselect three preview images
2023-02-22 09:15:03 +00:00
idx_img_prev = np.round(np.linspace(0, len(imgs_transition) - 1, 5)[1:-1]).astype(np.int32)
2023-01-15 11:02:11 +00:00
list_imgs_preview = []
2023-02-15 17:21:00 +00:00
for j in idx_img_prev:
2023-01-15 11:02:11 +00:00
list_imgs_preview.append(Image.fromarray(imgs_transition[j]))
2023-02-22 09:15:03 +00:00
2023-01-15 15:52:42 +00:00
# Save the preview imgs as jpgs on disk so we are not sending umcompressed data around
2023-02-22 09:15:03 +00:00
current_timestamp = get_time('second')
2023-01-15 15:52:42 +00:00
self.list_fp_imgs_current = []
2023-01-15 11:02:11 +00:00
for i in range(len(list_imgs_preview)):
2023-02-22 09:15:03 +00:00
fp_img = os.path.join(self.dp_imgs, f"img_preview_{i}_{current_timestamp}.jpg")
2023-01-15 11:02:11 +00:00
list_imgs_preview[i].save(fp_img)
2023-01-15 15:52:42 +00:00
self.list_fp_imgs_current.append(fp_img)
# Insert cheap frames for the movie
2023-02-15 17:21:00 +00:00
imgs_transition_ext = add_frames_linear_interp(imgs_transition, self.duration_video, self.fps)
2023-01-08 09:33:45 +00:00
2023-01-15 11:02:11 +00:00
# Save as movie
self.fp_movie = self.get_fp_video_last()
2023-02-18 06:56:30 +00:00
if os.path.isfile(self.fp_movie):
os.remove(self.fp_movie)
ms = MovieSaver(self.fp_movie, fps=self.fps)
2023-01-08 09:33:45 +00:00
for img in tqdm(imgs_transition_ext):
ms.write_frame(img)
ms.finalize()
2023-01-15 11:02:11 +00:00
print("DONE SAVING MOVIE! SENDING BACK...")
2023-02-22 09:15:03 +00:00
2023-01-15 15:52:42 +00:00
# Assemble Output, updating the preview images and le movie
2023-02-18 06:56:30 +00:00
list_return = self.list_fp_imgs_current + [self.fp_movie]
2023-01-15 11:02:11 +00:00
return list_return
2023-01-15 15:52:42 +00:00
def stack_forward(self, prompt2, seed2):
2023-02-22 09:15:03 +00:00
r"""
Allows to generate multi-segment movies. Sets last image -> first image with all
relevant parameters.
"""
2023-01-15 15:52:42 +00:00
# Save preview images, prompts and seeds into dictionary for stacking
2023-02-20 07:29:21 +00:00
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)
2023-02-22 09:15:03 +00:00
idx_segment = len(self.list_all_segments)
2023-02-20 07:29:21 +00:00
dp_segment = os.path.join(self.dp_session, f"segment_{str(idx_segment).zfill(3)}")
2023-02-22 09:15:03 +00:00
2023-02-20 07:29:21 +00:00
self.list_all_segments.append(dp_segment)
self.lb.write_imgs_transition(dp_segment)
2023-02-22 09:15:03 +00:00
fp_movie_last = self.get_fp_video_last()
fp_movie_next = self.get_fp_video_next()
2023-02-22 09:15:03 +00:00
shutil.copyfile(fp_movie_last, fp_movie_next)
2023-02-22 09:15:03 +00:00
2023-02-21 18:53:26 +00:00
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"))
2023-01-15 15:52:42 +00:00
self.lb.swap_forward()
2023-02-22 09:15:03 +00:00
2023-02-21 18:53:26 +00:00
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"))
2023-02-19 14:46:02 +00:00
fp_multi = self.multi_concat()
list_out = [fp_multi]
2023-02-22 09:15:03 +00:00
list_out.extend([os.path.join(self.dp_imgs, f"img2_{self.user_id}.jpg")])
2023-02-22 09:15:03 +00:00
list_out.extend([self.fp_img_empty] * 4)
2023-02-19 14:46:02 +00:00
list_out.append(gr.update(interactive=False, value=prompt2))
list_out.append(gr.update(interactive=False, value=seed2))
2023-01-15 15:52:42 +00:00
list_out.append("")
list_out.append(np.random.randint(0, 10000000))
2023-02-19 14:46:02 +00:00
print(f"stack_forward: fp_multi {fp_multi}")
return list_out
2023-02-18 06:56:30 +00:00
def multi_concat(self):
2023-02-22 09:15:03 +00:00
r"""
Concatentates all stacked segments into one long movie.
"""
list_fp_movies = self.get_fp_video_all()
2023-02-18 06:56:30 +00:00
# Concatenate movies and save
fp_final = os.path.join(self.dp_session, f"concat_{self.user_id}.mp4")
2023-02-18 06:56:30 +00:00
concatenate_movies(fp_final, list_fp_movies)
return fp_final
2023-02-19 14:46:02 +00:00
def get_fp_video_all(self):
2023-02-22 09:15:03 +00:00
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
2023-02-22 09:15:03 +00:00
def get_fp_video_next(self):
2023-02-22 09:15:03 +00:00
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
2023-02-22 09:15:03 +00:00
def get_fp_video_last(self):
2023-02-22 09:15:03 +00:00
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
2023-02-22 09:15:03 +00:00
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)
2023-01-12 03:11:56 +00:00
with gr.Blocks() as demo:
2023-02-23 09:32:53 +00:00
gr.HTML("""<h1>Latent Blending</h1>
<p>Create butter-smooth transitions between prompts, powered by stable diffusion</p>
2023-02-22 17:36:14 +00:00
<p>For faster inference without waiting in queue, you may duplicate the space and upgrade to GPU in settings.
<br/>
2023-02-23 09:32:53 +00:00
<a href="https://huggingface.co/spaces/lunarring/latentblending?duplicate=true">
2023-02-22 17:36:14 +00:00
<img style="margin-top: 0em; margin-bottom: 0em" src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>
2023-02-23 09:32:53 +00:00
</p>""")
2023-02-22 17:36:14 +00:00
2023-02-21 11:15:55 +00:00
with gr.Row():
prompt1 = gr.Textbox(label="prompt 1")
prompt2 = gr.Textbox(label="prompt 2")
2023-02-22 09:15:03 +00:00
2023-02-21 11:15:55 +00:00
with gr.Row():
2023-02-22 17:36:14 +00:00
duration_compute = gr.Slider(10, 25, bf.t_compute_max_allowed, step=1, label='waiting time', interactive=True)
2023-02-22 09:15:03 +00:00
duration_video = gr.Slider(1, 100, bf.duration_video, step=0.1, label='video duration', interactive=True)
2023-02-21 11:15:55 +00:00
height = gr.Slider(256, 2048, bf.height, step=128, label='height', interactive=True)
2023-02-22 09:15:03 +00:00
width = gr.Slider(256, 2048, bf.width, step=128, label='width', interactive=True)
2023-02-21 11:15:55 +00:00
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)
2023-02-22 09:15:03 +00:00
guidance_scale = gr.Slider(1, 25, bf.guidance_scale, step=0.1, label='guidance_scale', interactive=True)
negative_prompt = gr.Textbox(label="negative prompt")
2023-02-21 11:15:55 +00:00
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')
2023-02-22 09:15:03 +00:00
2023-02-21 11:15:55 +00:00
with gr.Accordion("Last image crossfeeding.", open=True):
with gr.Row():
2023-02-22 09:15:03 +00:00
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)
2023-02-21 11:15:55 +00:00
with gr.Accordion("Transition settings", open=True):
with gr.Row():
2023-02-22 09:15:03 +00:00
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)
2023-02-21 11:15:55 +00:00
with gr.Row():
2023-02-22 09:15:03 +00:00
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)
2023-02-21 11:15:55 +00:00
with gr.Row():
2023-02-22 17:36:14 +00:00
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')
2023-02-22 09:15:03 +00:00
2023-02-21 11:15:55 +00:00
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")
2023-02-22 09:15:03 +00:00
2023-02-21 11:15:55 +00:00
with gr.Row():
2023-02-22 09:15:03 +00:00
vid_single = gr.Video(label="current single trans")
vid_multi = gr.Video(label="concatented multi trans")
2023-02-21 11:15:55 +00:00
with gr.Row():
b_stackforward = gr.Button('append last movie segment (left) to multi movie (right)', variant='primary')
2023-02-22 09:15:03 +00:00
2023-02-21 11:52:57 +00:00
with gr.Row():
gr.Markdown(
"""
# Parameters
## Main
- compute budget: set your waiting time for the transition. high values = better quality
- video duration: seconds per segment
- height/width: in pixels
2023-02-22 09:15:03 +00:00
2023-02-21 11:52:57 +00:00
## 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
2023-02-22 09:15:03 +00:00
2023-02-21 11:54:34 +00:00
## Last image crossfeeding
2023-02-21 11:52:57 +00:00
- 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.
2023-02-22 09:15:03 +00:00
2023-02-21 11:52:57 +00:00
## 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.
2023-02-22 09:15:03 +00:00
""")
with gr.Row():
user_id = gr.Textbox(label="user id", interactive=False)
2023-02-22 09:15:03 +00:00
2023-02-21 11:15:55 +00:00
# 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
2023-02-22 09:15:03 +00:00
2023-02-21 11:15:55 +00:00
dict_ui_elem["duration_compute"] = duration_compute
dict_ui_elem["duration_video"] = duration_video
dict_ui_elem["height"] = height
dict_ui_elem["width"] = width
2023-02-22 09:15:03 +00:00
2023-02-21 11:15:55 +00:00
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
2023-02-22 09:15:03 +00:00
2023-02-21 11:15:55 +00:00
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
2023-02-22 09:15:03 +00:00
2023-02-21 11:15:55 +00:00
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
2023-02-22 09:15:03 +00:00
2023-02-21 11:15:55 +00:00
# Convert to list, as gradio doesn't seem to accept dicts
2023-02-22 09:15:03 +00:00
list_ui_vals = []
2023-02-21 11:15:55 +00:00
list_ui_keys = []
for k in dict_ui_elem.keys():
2023-02-22 09:15:03 +00:00
list_ui_vals.append(dict_ui_elem[k])
2023-02-21 11:15:55 +00:00
list_ui_keys.append(k)
bf.list_ui_keys = list_ui_keys
2023-02-22 09:15:03 +00:00
2023-02-21 11:15:55 +00:00
b_newseed1.click(bf.randomize_seed1, outputs=seed1)
b_newseed2.click(bf.randomize_seed2, outputs=seed2)
2023-02-22 09:15:03 +00:00
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])
2023-02-20 07:29:21 +00:00
demo.launch(share=bf.share, inbrowser=True, inline=False)