revert back to 3s

app.py

@@ -64,14 +64,14 @@ def load_model():
 model = load_model()
 
 # Text-to-video generation function
-@spaces.GPU(duration=200)
+@spaces.GPU(duration=120)
 def generate_video(prompt, image=None, duration=5, guidance_scale=9, video_guidance_scale=5, progress=gr.Progress(track_tqdm=True)):
-    multiplier = 3
-    temp = int(duration * multiplier) + 1 
+    multiplier = 0.8 if is_canonical else 2.4
+    temp = int(duration * 0.8)  # Convert seconds to temp value (assuming 24 FPS)
     torch_dtype = torch.bfloat16 if MODEL_DTYPE == "bf16" else torch.float32
     if(image):
-        cropped_image = center_crop(image, 1280, 768)
-        resized_image = cropped_image.resize((1280, 768))
+        cropped_image = center_crop(image, 1280, 720)
+        resized_image = cropped_image.resize((1280, 720))
         with torch.no_grad(), torch.cuda.amp.autocast(enabled=True, dtype=torch_dtype):
             frames = model.generate_i2v(
                 prompt=prompt,
@@ -97,26 +97,23 @@ def generate_video(prompt, image=None, duration=5, guidance_scale=9, video_guida
                 output_type="pil",
                 save_memory=True,
             )
-    return frames, gr.update()
-
-def compose_video(frames):
     output_path = f"{str(uuid.uuid4())}_output_video.mp4"
-    export_to_video(frames, output_path, fps=24)
-    return output_path 
+    export_to_video(frames, output_path, fps=8 if is_canonical else 24)
+    return output_path
 
 # Gradio interface
 with gr.Blocks() as demo:
     gr.Markdown("# Pyramid Flow")
     gr.Markdown("Pyramid Flow is a training-efficient Autoregressive Video Generation model based on Flow Matching. It is trained only on open-source datasets within 20.7k A100 GPU hours")
     gr.Markdown("[[Paper](https://arxiv.org/pdf/2410.05954)], [[Model](https://huggingface.co/rain1011/pyramid-flow-sd3)], [[Code](https://github.com/jy0205/Pyramid-Flow)]")
-    frames = gr.State()
+    
     with gr.Row():
         with gr.Column():
             with gr.Accordion("Image to Video (optional)", open=False):
                 i2v_image = gr.Image(type="pil", label="Input Image")
             t2v_prompt = gr.Textbox(label="Prompt")
             with gr.Accordion("Advanced settings", open=False):
-                t2v_duration = gr.Slider(minimum=1, maximum=2 if is_canonical else 10, value=2 if is_canonical else 5, step=1, label="Duration (seconds)")
+                t2v_duration = gr.Slider(minimum=1, maximum=10, value=5, step=1, label="Duration (seconds)", visible=not is_canonical)
                 t2v_guidance_scale = gr.Slider(minimum=1, maximum=15, value=9, step=0.1, label="Guidance Scale")
                 t2v_video_guidance_scale = gr.Slider(minimum=1, maximum=15, value=5, step=0.1, label="Video Guidance Scale")
             t2v_generate_btn = gr.Button("Generate Video")
@@ -145,10 +142,6 @@ with gr.Blocks() as demo:
     t2v_generate_btn.click(
         generate_video,
         inputs=[t2v_prompt, i2v_image, t2v_duration, t2v_guidance_scale, t2v_video_guidance_scale],
-        outputs=[frames, t2v_output]
-    ).then(
-        compose_video,
-        inputs=[frames],
         outputs=t2v_output
     )