Update app.py
Browse files
app.py
CHANGED
@@ -34,6 +34,10 @@ dtype = torch.float16
|
|
34 |
pipe = AnimateDiffPipeline.from_pretrained(bases[base_loaded], torch_dtype=dtype).to(device)
|
35 |
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing", beta_schedule="linear")
|
36 |
|
|
|
|
|
|
|
|
|
37 |
# unfortunately 2 steps isn't good enough for AiTube, we need 4 steps
|
38 |
step = 4
|
39 |
repo = "ByteDance/AnimateDiff-Lightning"
|
@@ -63,7 +67,7 @@ step_loaded = step
|
|
63 |
# This is a critical issue for AiTube so we are forced to implement our own routine.
|
64 |
# ------------------------------------------------------------------------------------
|
65 |
|
66 |
-
def export_to_video_file(video_frames, output_video_path=None, fps=
|
67 |
if output_video_path is None:
|
68 |
output_video_path = tempfile.NamedTemporaryFile(suffix=".webm").name
|
69 |
|
@@ -92,7 +96,7 @@ def export_to_video_file(video_frames, output_video_path=None, fps=10):
|
|
92 |
# those are way too slow for a AiTube which needs things to be as fast as possible
|
93 |
# -----------------------------------------------------------------------------------
|
94 |
|
95 |
-
def interpolate_video_frames(input_file_path, output_file_path, output_fps=
|
96 |
scale_factor = original_duration / desired_duration
|
97 |
interpolation_filter = f'minterpolate=fps={output_fps},setpts={scale_factor}*PTS'
|
98 |
|
@@ -165,12 +169,12 @@ def generate_image(secret_token, prompt, base, width, height, motion, step, desi
|
|
165 |
#
|
166 |
# maybe to make things faster, we could *not* encode the video (as this uses files and external processes, which can be slow)
|
167 |
# and instead return the unencoded frames to the frontend renderer?
|
168 |
-
raw_video_path = export_to_video_file(output.frames[0], raw_video_path, fps=
|
169 |
|
170 |
final_video_path = raw_video_path
|
171 |
|
172 |
# Optional frame interpolation
|
173 |
-
if desired_duration
|
174 |
final_video_path = interpolate_video_frames(raw_video_path, enhanced_video_path, output_fps=desired_fps, desired_duration=desired_duration)
|
175 |
|
176 |
# Read the content of the video file and encode it to base64
|
@@ -258,8 +262,8 @@ with gr.Blocks() as demo:
|
|
258 |
('8-Step', 8)],
|
259 |
value=4,
|
260 |
)
|
261 |
-
duration_slider = gr.Slider(label="Desired Duration (seconds)", min_value=1, max_value=120, value=
|
262 |
-
fps_slider = gr.Slider(label="Desired Frames Per Second", min_value=10, max_value=60, value=
|
263 |
|
264 |
submit = gr.Button()
|
265 |
|
|
|
34 |
pipe = AnimateDiffPipeline.from_pretrained(bases[base_loaded], torch_dtype=dtype).to(device)
|
35 |
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing", beta_schedule="linear")
|
36 |
|
37 |
+
# those are AnimateDiff defaults - we don't touch them for now
|
38 |
+
hardcoded_fps = 10
|
39 |
+
hardcoded_duration_sec = 1.6
|
40 |
+
|
41 |
# unfortunately 2 steps isn't good enough for AiTube, we need 4 steps
|
42 |
step = 4
|
43 |
repo = "ByteDance/AnimateDiff-Lightning"
|
|
|
67 |
# This is a critical issue for AiTube so we are forced to implement our own routine.
|
68 |
# ------------------------------------------------------------------------------------
|
69 |
|
70 |
+
def export_to_video_file(video_frames, output_video_path=None, fps=hardcoded_fps):
|
71 |
if output_video_path is None:
|
72 |
output_video_path = tempfile.NamedTemporaryFile(suffix=".webm").name
|
73 |
|
|
|
96 |
# those are way too slow for a AiTube which needs things to be as fast as possible
|
97 |
# -----------------------------------------------------------------------------------
|
98 |
|
99 |
+
def interpolate_video_frames(input_file_path, output_file_path, output_fps=hardcoded_fps, desired_duration=hardcoded_duration_sec, original_duration=hardcoded_duration_sec):
|
100 |
scale_factor = original_duration / desired_duration
|
101 |
interpolation_filter = f'minterpolate=fps={output_fps},setpts={scale_factor}*PTS'
|
102 |
|
|
|
169 |
#
|
170 |
# maybe to make things faster, we could *not* encode the video (as this uses files and external processes, which can be slow)
|
171 |
# and instead return the unencoded frames to the frontend renderer?
|
172 |
+
raw_video_path = export_to_video_file(output.frames[0], raw_video_path, fps=hardcoded_fps)
|
173 |
|
174 |
final_video_path = raw_video_path
|
175 |
|
176 |
# Optional frame interpolation
|
177 |
+
if desired_duration > hardcoded_duration_sec or desired_duration < hardcoded_duration_sec or desired_fps > hardcoded_fps or desired_fps < hardcoded_fps:
|
178 |
final_video_path = interpolate_video_frames(raw_video_path, enhanced_video_path, output_fps=desired_fps, desired_duration=desired_duration)
|
179 |
|
180 |
# Read the content of the video file and encode it to base64
|
|
|
262 |
('8-Step', 8)],
|
263 |
value=4,
|
264 |
)
|
265 |
+
duration_slider = gr.Slider(label="Desired Duration (seconds)", min_value=1, max_value=120, value=hardcoded_duration_sec, step=0.1)
|
266 |
+
fps_slider = gr.Slider(label="Desired Frames Per Second", min_value=10, max_value=60, value=hardcoded_fps, step=1)
|
267 |
|
268 |
submit = gr.Button()
|
269 |
|