Try with 121 frames + touch-ups

app.py

@@ -25,7 +25,7 @@ MAX_SEED = np.iinfo(np.int32).max
 
 FIXED_FPS = 24
 MIN_FRAMES_MODEL = 8
-MAX_FRAMES_MODEL = 81
+MAX_FRAMES_MODEL = 121
 
 MIN_DURATION = round(MIN_FRAMES_MODEL/FIXED_FPS,1)
 MAX_DURATION = round(MAX_FRAMES_MODEL/FIXED_FPS,1)
@@ -56,7 +56,7 @@ optimize_pipeline_(pipe,
 
 
 default_prompt_i2v = "make this image come alive, cinematic motion, smooth animation"
-default_negative_prompt = "色调艳丽, 过曝, 静态, 细节模糊不清, 字幕, 风格, 作品, 画作, 画面, 静止, 整体发灰, 最差质量, 低质量, JPEG压缩残留, 丑陋的, 残缺的, 多余的手指, 画得不好的手部, 画得不好的脸部, 畸形的, 毁容的, 形态畸形的肢体, 手指融合, 静止不动的画面, 杂乱的背景, 三条腿, 背景人很多, 倒着走"
+default_negative_prompt = "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走"
 
 
 def resize_image(image: Image.Image) -> Image.Image:
@@ -100,8 +100,8 @@ def generate_video(
     prompt,
     negative_prompt=default_negative_prompt,
     duration_seconds = MAX_DURATION,
-    guidance_scale = 1,
-    steps = 4,
+    guidance_scale = 3.5,
+    steps = 28,
     seed = 42,
     randomize_seed = False,
     progress=gr.Progress(track_tqdm=True),
@@ -184,7 +184,7 @@ with gr.Blocks() as demo:
                 negative_prompt_input = gr.Textbox(label="Negative Prompt", value=default_negative_prompt, lines=3)
                 seed_input = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=42, interactive=True)
                 randomize_seed_checkbox = gr.Checkbox(label="Randomize seed", value=True, interactive=True)
-                steps_slider = gr.Slider(minimum=1, maximum=30, step=1, value=4, label="Inference Steps") 
+                steps_slider = gr.Slider(minimum=1, maximum=40, step=1, value=28, label="Inference Steps")
                 guidance_scale_input = gr.Slider(minimum=0.0, maximum=20.0, step=0.5, value=1.0, label="Guidance Scale", visible=False)
 
             generate_button = gr.Button("Generate Video", variant="primary")

optimization.py

@@ -20,7 +20,7 @@ from optimization_utils import ZeroGPUCompiledModel
 P = ParamSpec('P')
 
 
-TRANSFORMER_NUM_FRAMES_DIM = torch.export.Dim('num_frames', min=3, max=21)
+TRANSFORMER_NUM_FRAMES_DIM = torch.export.Dim('num_frames', min=3, max=31)
 
 TRANSFORMER_DYNAMIC_SHAPES = {
     'hidden_states': {