Spaces:
Sleeping
Sleeping
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,245 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# -*- coding: utf-8 -*-
|
2 |
+
import os
|
3 |
+
import inspect
|
4 |
+
import torch
|
5 |
+
from diffusers import StableDiffusionPipeline
|
6 |
+
from PIL import Image
|
7 |
+
import numpy as np
|
8 |
+
from torch import autocast
|
9 |
+
import cv2
|
10 |
+
import gradio as gr
|
11 |
+
|
12 |
+
# -----------------------------------------------------------------------------
|
13 |
+
# 1. REQUIREMENTS & SETUP
|
14 |
+
# -----------------------------------------------------------------------------
|
15 |
+
# To set up the environment for this script, create a file named 'requirements.txt'
|
16 |
+
# with the following content and run 'pip install -r requirements.txt':
|
17 |
+
#
|
18 |
+
# torch>=2.0.0
|
19 |
+
# torchvision>=0.15.1
|
20 |
+
# diffusers>=0.20.2
|
21 |
+
# transformers>=4.30.2
|
22 |
+
# accelerate>=0.21.0
|
23 |
+
# gradio>=3.36.1
|
24 |
+
# opencv-python-headless>=4.8.0.74
|
25 |
+
# -----------------------------------------------------------------------------
|
26 |
+
|
27 |
+
# --- Automatic Device Detection ---
|
28 |
+
torch_device = "cuda" if torch.cuda.is_available() else "cpu"
|
29 |
+
print("-------------------------------------------------")
|
30 |
+
print(f"INFO: Using device: {torch_device.upper()}")
|
31 |
+
if torch_device == "cpu":
|
32 |
+
print("WARNING: CUDA (GPU) not detected. The script will run on the CPU.")
|
33 |
+
print(" This will be extremely slow. For better performance,")
|
34 |
+
print(" please ensure you have an NVIDIA GPU and the correct")
|
35 |
+
print(" PyTorch version with CUDA support installed.")
|
36 |
+
print("-------------------------------------------------")
|
37 |
+
|
38 |
+
|
39 |
+
# --- Load the Model ---
|
40 |
+
print("Loading Stable Diffusion model... This may take a moment.")
|
41 |
+
try:
|
42 |
+
# Load the pipeline and move it to the detected device
|
43 |
+
pipe = StableDiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-2-1-base")
|
44 |
+
pipe.to(torch_device)
|
45 |
+
print("Model loaded successfully.")
|
46 |
+
except Exception as e:
|
47 |
+
print(f"Error loading model: {e}")
|
48 |
+
print("Please check your internet connection and ensure the model name is correct.")
|
49 |
+
exit()
|
50 |
+
|
51 |
+
# -----------------------------------------------------------------------------
|
52 |
+
# Helper Functions (slerp, diffuse)
|
53 |
+
# -----------------------------------------------------------------------------
|
54 |
+
|
55 |
+
@torch.no_grad()
|
56 |
+
def diffuse(
|
57 |
+
pipe, cond_embeddings, cond_latents, num_inference_steps, guidance_scale, eta
|
58 |
+
):
|
59 |
+
# This function remains the same, as it gets the device from the input tensors
|
60 |
+
device = cond_latents.get_device()
|
61 |
+
max_length = cond_embeddings.shape[1]
|
62 |
+
uncond_input = pipe.tokenizer([""], padding="max_length", max_length=max_length, return_tensors="pt")
|
63 |
+
uncond_embeddings = pipe.text_encoder(uncond_input.input_ids.to(device))[0]
|
64 |
+
text_embeddings = torch.cat([uncond_embeddings, cond_embeddings])
|
65 |
+
|
66 |
+
if "LMS" in pipe.scheduler.__class__.__name__:
|
67 |
+
cond_latents = cond_latents * pipe.scheduler.sigmas[0]
|
68 |
+
|
69 |
+
accepts_offset = "offset" in set(inspect.signature(pipe.scheduler.set_timesteps).parameters.keys())
|
70 |
+
extra_set_kwargs = {}
|
71 |
+
if accepts_offset:
|
72 |
+
extra_set_kwargs["offset"] = 1
|
73 |
+
pipe.scheduler.set_timesteps(num_inference_steps, **extra_set_kwargs)
|
74 |
+
|
75 |
+
accepts_eta = "eta" in set(inspect.signature(pipe.scheduler.step).parameters.keys())
|
76 |
+
extra_step_kwargs = {}
|
77 |
+
if accepts_eta:
|
78 |
+
extra_step_kwargs["eta"] = eta
|
79 |
+
|
80 |
+
for i, t in enumerate(pipe.scheduler.timesteps):
|
81 |
+
latent_model_input = torch.cat([cond_latents] * 2)
|
82 |
+
if "LMS" in pipe.scheduler.__class__.__name__:
|
83 |
+
sigma = pipe.scheduler.sigmas[i]
|
84 |
+
latent_model_input = latent_model_input / ((sigma**2 + 1) ** 0.5)
|
85 |
+
|
86 |
+
# predict the noise residual
|
87 |
+
noise_pred = pipe.unet(latent_model_input, t, encoder_hidden_states=text_embeddings)["sample"]
|
88 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
89 |
+
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
|
90 |
+
cond_latents = pipe.scheduler.step(noise_pred, t, cond_latents, **extra_step_kwargs)["prev_sample"]
|
91 |
+
|
92 |
+
cond_latents = 1 / 0.18215 * cond_latents
|
93 |
+
image = pipe.vae.decode(cond_latents).sample
|
94 |
+
image = (image / 2 + 0.5).clamp(0, 1)
|
95 |
+
image = image.cpu().permute(0, 2, 3, 1).numpy()
|
96 |
+
image = (image[0] * 255).astype(np.uint8)
|
97 |
+
return image
|
98 |
+
|
99 |
+
def slerp(t, v0, v1, DOT_THRESHOLD=0.9995):
|
100 |
+
# This function is device-agnostic
|
101 |
+
inputs_are_torch = isinstance(v0, torch.Tensor)
|
102 |
+
if inputs_are_torch:
|
103 |
+
input_device = v0.device
|
104 |
+
v0 = v0.cpu().numpy()
|
105 |
+
v1 = v1.cpu().numpy()
|
106 |
+
dot = np.sum(v0 * v1 / (np.linalg.norm(v0) * np.linalg.norm(v1)))
|
107 |
+
if np.abs(dot) > DOT_THRESHOLD:
|
108 |
+
v2 = (1 - t) * v0 + t * v1
|
109 |
+
else:
|
110 |
+
theta_0 = np.arccos(dot)
|
111 |
+
sin_theta_0 = np.sin(theta_0)
|
112 |
+
theta_t = theta_0 * t
|
113 |
+
sin_theta_t = np.sin(theta_t)
|
114 |
+
s0 = np.sin(theta_0 - theta_t) / sin_theta_0
|
115 |
+
s1 = sin_theta_t / sin_theta_0
|
116 |
+
v2 = s0 * v0 + s1 * v1
|
117 |
+
if inputs_are_torch:
|
118 |
+
v2 = torch.from_numpy(v2).to(input_device)
|
119 |
+
return v2
|
120 |
+
|
121 |
+
# -----------------------------------------------------------------------------
|
122 |
+
# Main Generator Function for Gradio
|
123 |
+
# -----------------------------------------------------------------------------
|
124 |
+
def generate_dream_video(
|
125 |
+
prompt_1, prompt_2, seed_1, seed_2,
|
126 |
+
width, height, num_steps, guidance_scale,
|
127 |
+
num_inference_steps, eta, name
|
128 |
+
):
|
129 |
+
# --- 1. SETUP ---
|
130 |
+
yield {
|
131 |
+
status_text: "Status: Preparing prompts and latents...",
|
132 |
+
live_frame: None,
|
133 |
+
output_video: None,
|
134 |
+
}
|
135 |
+
prompts = [prompt_1, prompt_2]
|
136 |
+
seeds = [int(seed_1), int(seed_2)]
|
137 |
+
rootdir = './dreams'
|
138 |
+
outdir = os.path.join(rootdir, name)
|
139 |
+
os.makedirs(outdir, exist_ok=True)
|
140 |
+
|
141 |
+
# --- 2. EMBEDDINGS AND LATENTS ---
|
142 |
+
prompt_embeddings = []
|
143 |
+
for prompt in prompts:
|
144 |
+
text_input = pipe.tokenizer(prompt, padding="max_length", max_length=pipe.tokenizer.model_max_length, truncation=True, return_tensors="pt")
|
145 |
+
# Move input_ids to the correct device before text encoding
|
146 |
+
with torch.no_grad():
|
147 |
+
embed = pipe.text_encoder(text_input.input_ids.to(torch_device))[0]
|
148 |
+
prompt_embeddings.append(embed)
|
149 |
+
|
150 |
+
prompt_embedding_a, prompt_embedding_b = prompt_embeddings
|
151 |
+
|
152 |
+
# Use a device-specific generator for reproducibility
|
153 |
+
generator_a = torch.Generator(device=torch_device).manual_seed(seeds[0])
|
154 |
+
generator_b = torch.Generator(device=torch_device).manual_seed(seeds[1])
|
155 |
+
|
156 |
+
init_a = torch.randn((1, pipe.unet.config.in_channels, height // 8, width // 8), device=torch_device, generator=generator_a)
|
157 |
+
init_b = torch.randn((1, pipe.unet.config.in_channels, height // 8, width // 8), device=torch_device, generator=generator_b)
|
158 |
+
|
159 |
+
# --- 3. GENERATION LOOP ---
|
160 |
+
frame_paths = []
|
161 |
+
for i, t in enumerate(np.linspace(0, 1, num_steps)):
|
162 |
+
yield {
|
163 |
+
status_text: f"Status: Generating frame {i + 1} of {num_steps} on {torch_device.upper()}...",
|
164 |
+
live_frame: None,
|
165 |
+
output_video: None,
|
166 |
+
}
|
167 |
+
|
168 |
+
cond_embedding = slerp(float(t), prompt_embedding_a, prompt_embedding_b)
|
169 |
+
init = slerp(float(t), init_a, init_b)
|
170 |
+
|
171 |
+
# Use autocast only if on CUDA
|
172 |
+
with autocast(torch_device) if torch_device == "cuda" else open(os.devnull, 'w') as f:
|
173 |
+
image = diffuse(pipe, cond_embedding, init, num_inference_steps, guidance_scale, eta)
|
174 |
+
|
175 |
+
im = Image.fromarray(image)
|
176 |
+
outpath = os.path.join(outdir, f'frame{i:06d}.jpg')
|
177 |
+
im.save(outpath)
|
178 |
+
frame_paths.append(outpath)
|
179 |
+
|
180 |
+
yield { live_frame: im }
|
181 |
+
|
182 |
+
# --- 4. VIDEO COMPILATION ---
|
183 |
+
yield { status_text: "Status: Compiling video from frames..." }
|
184 |
+
|
185 |
+
video_path = os.path.join(outdir, f"{name}.mp4")
|
186 |
+
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
|
187 |
+
video_writer = cv2.VideoWriter(video_path, fourcc, 15, (width, height))
|
188 |
+
for frame_path in frame_paths:
|
189 |
+
frame = cv2.imread(frame_path)
|
190 |
+
video_writer.write(frame)
|
191 |
+
video_writer.release()
|
192 |
+
|
193 |
+
print(f"Video saved to {video_path}")
|
194 |
+
yield {
|
195 |
+
status_text: f"Status: Done! Video saved to {video_path}",
|
196 |
+
output_video: video_path
|
197 |
+
}
|
198 |
+
|
199 |
+
# -----------------------------------------------------------------------------
|
200 |
+
# Gradio UI (Unchanged)
|
201 |
+
# -----------------------------------------------------------------------------
|
202 |
+
with gr.Blocks(theme=gr.themes.Soft(), css="footer {display: none !important}") as demo:
|
203 |
+
gr.Markdown("# 🎥 Stable Diffusion Video Interpolation")
|
204 |
+
gr.Markdown("Create smooth transition videos between two concepts. Configure the prompts and settings below, then click Generate.")
|
205 |
+
|
206 |
+
with gr.Row():
|
207 |
+
with gr.Column(scale=2):
|
208 |
+
with gr.Accordion("1. Core Prompts & Seeds", open=True):
|
209 |
+
prompt_1 = gr.Textbox(lines=2, label="Starting Prompt", value="ultrarealistic steam punk neural network machine in the shape of a brain, placed on a pedestal, covered with neurons made of gears.")
|
210 |
+
seed_1 = gr.Number(label="Seed 1", value=243, precision=0, info="A specific number to control the starting noise pattern.")
|
211 |
+
prompt_2 = gr.Textbox(lines=2, label="Ending Prompt", value="A bioluminescent, glowing jellyfish floating in a dark, deep abyss, surrounded by sparkling plankton.")
|
212 |
+
seed_2 = gr.Number(label="Seed 2", value=523, precision=0, info="A specific number to control the ending noise pattern.")
|
213 |
+
name = gr.Textbox(label="Output File Name", value="my_dream_video", info="The name for the output folder and .mp4 file.")
|
214 |
+
|
215 |
+
with gr.Accordion("2. Generation Parameters", open=True):
|
216 |
+
with gr.Row():
|
217 |
+
width = gr.Slider(label="Width", minimum=256, maximum=1024, value=512, step=64)
|
218 |
+
height = gr.Slider(label="Height", minimum=256, maximum=1024, value=512, step=64)
|
219 |
+
num_steps = gr.Slider(label="Interpolation Frames", minimum=10, maximum=500, value=120, step=1, info="How many frames the final video will have. More frames = smoother video.")
|
220 |
+
|
221 |
+
with gr.Accordion("3. Advanced Diffusion Settings", open=False):
|
222 |
+
num_inference_steps = gr.Slider(label="Inference Steps per Frame", minimum=10, maximum=100, value=40, step=1, info="More steps can improve quality but will be much slower.")
|
223 |
+
guidance_scale = gr.Slider(label="Guidance Scale (CFG)", minimum=1, maximum=20, value=7.5, step=0.5, info="How strongly the prompt guides the image generation.")
|
224 |
+
eta = gr.Slider(label="ETA (for DDIM Scheduler)", minimum=0.0, maximum=1.0, value=0.0, step=0.1, info="A parameter for noise scheduling. 0.0 is deterministic.")
|
225 |
+
|
226 |
+
run_button = gr.Button("Generate Video", variant="primary")
|
227 |
+
|
228 |
+
with gr.Column(scale=3):
|
229 |
+
status_text = gr.Textbox(label="Status", value="Ready", interactive=False)
|
230 |
+
live_frame = gr.Image(label="Live Preview", type="pil")
|
231 |
+
output_video = gr.Video(label="Final Video")
|
232 |
+
|
233 |
+
run_button.click(
|
234 |
+
fn=generate_dream_video,
|
235 |
+
inputs=[
|
236 |
+
prompt_1, prompt_2, seed_1, seed_2,
|
237 |
+
width, height, num_steps, guidance_scale,
|
238 |
+
num_inference_steps, eta, name
|
239 |
+
],
|
240 |
+
outputs=[status_text, live_frame, output_video]
|
241 |
+
)
|
242 |
+
|
243 |
+
# --- Launch the App ---
|
244 |
+
if __name__ == "__main__":
|
245 |
+
demo.launch(share=True, debug=True)
|