Spaces:
Paused
Paused
Update app.py
Browse files
app.py
CHANGED
@@ -1,73 +1,51 @@
|
|
1 |
import spaces
|
2 |
import gradio as gr
|
3 |
import torch
|
|
|
|
|
|
|
4 |
from diffusers import DiffusionPipeline
|
5 |
from diffusers.quantizers import PipelineQuantizationConfig
|
6 |
-
import imageio
|
7 |
from diffusers.utils.export_utils import export_to_video
|
8 |
-
import gradio as gr
|
9 |
-
import tempfile
|
10 |
|
11 |
-
|
12 |
-
import re
|
13 |
-
import json
|
14 |
-
import random
|
15 |
-
import tempfile
|
16 |
-
import traceback
|
17 |
-
from functools import partial
|
18 |
-
import numpy as np
|
19 |
-
from PIL import Image
|
20 |
-
import random
|
21 |
-
import numpy as np
|
22 |
-
import random
|
23 |
-
import gradio as gr
|
24 |
-
import tempfile
|
25 |
-
import numpy as np
|
26 |
-
from PIL import Image
|
27 |
-
import random
|
28 |
LANDSCAPE_WIDTH = 832
|
29 |
LANDSCAPE_HEIGHT = 480
|
30 |
MAX_SEED = np.iinfo(np.int32).max
|
31 |
-
|
32 |
FIXED_FPS = 16
|
33 |
MIN_FRAMES_MODEL = 8
|
34 |
MAX_FRAMES_MODEL = 81
|
35 |
T2V_FIXED_FPS = 16
|
36 |
-
MIN_DURATION = round(MIN_FRAMES_MODEL/FIXED_FPS,1)
|
37 |
-
MAX_DURATION = round(MAX_FRAMES_MODEL/FIXED_FPS,1)
|
|
|
38 |
# Checkpoint ID
|
39 |
ckpt_id = "Wan-AI/Wan2.1-T2V-14B-Diffusers"
|
40 |
|
41 |
-
#
|
42 |
quant_config = PipelineQuantizationConfig(
|
43 |
quant_backend="bitsandbytes_4bit",
|
44 |
quant_kwargs={
|
45 |
"load_in_4bit": True,
|
46 |
"bnb_4bit_quant_type": "nf4",
|
47 |
-
"bnb_4bit_compute_dtype": torch.bfloat16
|
48 |
},
|
49 |
-
components_to_quantize=["transformer", "text_encoder"]
|
50 |
)
|
51 |
|
52 |
-
# Load pipeline
|
53 |
pipe = DiffusionPipeline.from_pretrained(
|
54 |
ckpt_id,
|
55 |
quantization_config=quant_config,
|
56 |
-
torch_dtype=torch.bfloat16
|
57 |
)
|
58 |
pipe.enable_model_cpu_offload()
|
59 |
-
torch._dynamo.config.recompile_limit = 1000
|
60 |
-
torch._dynamo.config.capture_dynamic_output_shape_ops = True
|
61 |
|
62 |
-
#
|
63 |
def get_duration(prompt, height, width, negative_prompt, duration_seconds, guidance_scale, steps, seed, randomize_seed, progress):
|
64 |
-
|
65 |
-
if duration_seconds <= 2.5:
|
66 |
-
return steps * 18
|
67 |
-
else:
|
68 |
-
return steps * 25
|
69 |
|
70 |
-
#
|
71 |
@spaces.GPU(duration=get_duration)
|
72 |
def generate_video(prompt, height, width, negative_prompt, duration_seconds,
|
73 |
guidance_scale, steps, seed, randomize_seed,
|
@@ -75,7 +53,7 @@ def generate_video(prompt, height, width, negative_prompt, duration_seconds,
|
|
75 |
|
76 |
num_frames = np.clip(int(round(duration_seconds * FIXED_FPS)),
|
77 |
MIN_FRAMES_MODEL, MAX_FRAMES_MODEL)
|
78 |
-
current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
|
79 |
|
80 |
output_frames_list = pipe(
|
81 |
prompt=prompt,
|
@@ -85,20 +63,17 @@ def generate_video(prompt, height, width, negative_prompt, duration_seconds,
|
|
85 |
num_frames=num_frames,
|
86 |
guidance_scale=float(guidance_scale),
|
87 |
num_inference_steps=int(steps),
|
88 |
-
generator=torch.
|
89 |
).frames[0]
|
90 |
|
91 |
-
filename = f"t2v_aaa.mp4"
|
92 |
temp_dir = tempfile.mkdtemp()
|
93 |
-
video_path = os.path.join(temp_dir,
|
94 |
export_to_video(output_frames_list, video_path, fps=T2V_FIXED_FPS)
|
95 |
|
96 |
print(f"✅ Video saved to: {video_path}")
|
97 |
-
|
98 |
-
return video_path, current_seed, gr.File(value=video_path, visible=True, label=download_label)
|
99 |
-
|
100 |
|
101 |
-
#
|
102 |
with gr.Blocks(css="body { max-width: 100vw; overflow-x: hidden; }") as demo:
|
103 |
gr.Markdown("## 🚀 Wan2.1 T2V - Text to Video Generator (Quantized, Smart Duration)")
|
104 |
with gr.Row():
|
@@ -116,8 +91,12 @@ with gr.Blocks(css="body { max-width: 100vw; overflow-x: hidden; }") as demo:
|
|
116 |
with gr.Column():
|
117 |
output_video = gr.Video(label="Generated Video")
|
118 |
|
119 |
-
ui_inputs = [
|
|
|
|
|
|
|
|
|
120 |
run_btn.click(fn=generate_video, inputs=ui_inputs, outputs=output_video)
|
121 |
|
122 |
-
# Launch
|
123 |
demo.launch()
|
|
|
1 |
import spaces
|
2 |
import gradio as gr
|
3 |
import torch
|
4 |
+
import numpy as np
|
5 |
+
import os
|
6 |
+
import tempfile
|
7 |
from diffusers import DiffusionPipeline
|
8 |
from diffusers.quantizers import PipelineQuantizationConfig
|
|
|
9 |
from diffusers.utils.export_utils import export_to_video
|
|
|
|
|
10 |
|
11 |
+
# Constants
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
12 |
LANDSCAPE_WIDTH = 832
|
13 |
LANDSCAPE_HEIGHT = 480
|
14 |
MAX_SEED = np.iinfo(np.int32).max
|
|
|
15 |
FIXED_FPS = 16
|
16 |
MIN_FRAMES_MODEL = 8
|
17 |
MAX_FRAMES_MODEL = 81
|
18 |
T2V_FIXED_FPS = 16
|
19 |
+
MIN_DURATION = round(MIN_FRAMES_MODEL / FIXED_FPS, 1)
|
20 |
+
MAX_DURATION = round(MAX_FRAMES_MODEL / FIXED_FPS, 1)
|
21 |
+
|
22 |
# Checkpoint ID
|
23 |
ckpt_id = "Wan-AI/Wan2.1-T2V-14B-Diffusers"
|
24 |
|
25 |
+
# Quantization config
|
26 |
quant_config = PipelineQuantizationConfig(
|
27 |
quant_backend="bitsandbytes_4bit",
|
28 |
quant_kwargs={
|
29 |
"load_in_4bit": True,
|
30 |
"bnb_4bit_quant_type": "nf4",
|
31 |
+
"bnb_4bit_compute_dtype": torch.bfloat16,
|
32 |
},
|
33 |
+
components_to_quantize=["transformer", "text_encoder"],
|
34 |
)
|
35 |
|
36 |
+
# Load pipeline
|
37 |
pipe = DiffusionPipeline.from_pretrained(
|
38 |
ckpt_id,
|
39 |
quantization_config=quant_config,
|
40 |
+
torch_dtype=torch.bfloat16,
|
41 |
)
|
42 |
pipe.enable_model_cpu_offload()
|
|
|
|
|
43 |
|
44 |
+
# Duration estimator
|
45 |
def get_duration(prompt, height, width, negative_prompt, duration_seconds, guidance_scale, steps, seed, randomize_seed, progress):
|
46 |
+
return steps * 18 if duration_seconds <= 2.5 else steps * 25
|
|
|
|
|
|
|
|
|
47 |
|
48 |
+
# Inference function
|
49 |
@spaces.GPU(duration=get_duration)
|
50 |
def generate_video(prompt, height, width, negative_prompt, duration_seconds,
|
51 |
guidance_scale, steps, seed, randomize_seed,
|
|
|
53 |
|
54 |
num_frames = np.clip(int(round(duration_seconds * FIXED_FPS)),
|
55 |
MIN_FRAMES_MODEL, MAX_FRAMES_MODEL)
|
56 |
+
current_seed = np.random.randint(0, MAX_SEED) if randomize_seed else int(seed)
|
57 |
|
58 |
output_frames_list = pipe(
|
59 |
prompt=prompt,
|
|
|
63 |
num_frames=num_frames,
|
64 |
guidance_scale=float(guidance_scale),
|
65 |
num_inference_steps=int(steps),
|
66 |
+
generator=torch.manual_seed(current_seed),
|
67 |
).frames[0]
|
68 |
|
|
|
69 |
temp_dir = tempfile.mkdtemp()
|
70 |
+
video_path = os.path.join(temp_dir, "t2v_output.mp4")
|
71 |
export_to_video(output_frames_list, video_path, fps=T2V_FIXED_FPS)
|
72 |
|
73 |
print(f"✅ Video saved to: {video_path}")
|
74 |
+
return video_path # Only return video
|
|
|
|
|
75 |
|
76 |
+
# Gradio UI
|
77 |
with gr.Blocks(css="body { max-width: 100vw; overflow-x: hidden; }") as demo:
|
78 |
gr.Markdown("## 🚀 Wan2.1 T2V - Text to Video Generator (Quantized, Smart Duration)")
|
79 |
with gr.Row():
|
|
|
91 |
with gr.Column():
|
92 |
output_video = gr.Video(label="Generated Video")
|
93 |
|
94 |
+
ui_inputs = [
|
95 |
+
prompt_input, height_input, width_input, negative_prompt_input,
|
96 |
+
duration_input, guidance_scale_input, steps_input, seed_input,
|
97 |
+
randomize_seed_checkbox
|
98 |
+
]
|
99 |
run_btn.click(fn=generate_video, inputs=ui_inputs, outputs=output_video)
|
100 |
|
101 |
+
# Launch
|
102 |
demo.launch()
|