Files changed (2) hide show
  1. app.py +53 -25
  2. optimization.py +21 -45
app.py CHANGED
@@ -19,8 +19,11 @@ from optimization import optimize_pipeline_
19
 
20
  MODEL_ID = "Wan-AI/Wan2.2-I2V-A14B-Diffusers"
21
 
22
- LANDSCAPE_WIDTH = 832
23
- LANDSCAPE_HEIGHT = 480
 
 
 
24
  MAX_SEED = np.iinfo(np.int32).max
25
 
26
  FIXED_FPS = 16
@@ -50,11 +53,14 @@ for i in range(3):
50
  torch.cuda.synchronize()
51
  torch.cuda.empty_cache()
52
 
 
 
 
53
  optimize_pipeline_(pipe,
54
- image=Image.new('RGB', (LANDSCAPE_WIDTH, LANDSCAPE_HEIGHT)),
55
  prompt='prompt',
56
- height=LANDSCAPE_HEIGHT,
57
- width=LANDSCAPE_WIDTH,
58
  num_frames=MAX_FRAMES_MODEL,
59
  )
60
 
@@ -62,28 +68,51 @@ optimize_pipeline_(pipe,
62
  default_prompt_i2v = "make this image come alive, cinematic motion, smooth animation"
63
  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"
64
 
65
-
66
  def resize_image(image: Image.Image) -> Image.Image:
67
- if image.height > image.width:
68
- transposed = image.transpose(Image.Transpose.ROTATE_90)
69
- resized = resize_image_landscape(transposed)
70
- return resized.transpose(Image.Transpose.ROTATE_270)
71
- return resize_image_landscape(image)
72
 
 
 
 
73
 
74
- def resize_image_landscape(image: Image.Image) -> Image.Image:
75
- target_aspect = LANDSCAPE_WIDTH / LANDSCAPE_HEIGHT
76
- width, height = image.size
77
- in_aspect = width / height
78
- if in_aspect > target_aspect:
79
- new_width = round(height * target_aspect)
80
- left = (width - new_width) // 2
81
- image = image.crop((left, 0, left + new_width, height))
 
 
 
 
 
 
 
 
 
 
 
82
  else:
83
- new_height = round(width / target_aspect)
84
- top = (height - new_height) // 2
85
- image = image.crop((0, top, width, top + new_height))
86
- return image.resize((LANDSCAPE_WIDTH, LANDSCAPE_HEIGHT), Image.LANCZOS)
 
 
 
 
 
 
 
 
 
 
 
87
 
88
  def get_duration(
89
  input_image,
@@ -147,7 +176,6 @@ def generate_video(
147
  gr.Error: If input_image is None (no image uploaded).
148
 
149
  Note:
150
- - The function automatically resizes the input image to the target dimensions
151
  - Frame count is calculated as duration_seconds * FIXED_FPS (24)
152
  - Output dimensions are adjusted to be multiples of MOD_VALUE (32)
153
  - The function uses GPU acceleration via the @spaces.GPU decorator
@@ -185,7 +213,7 @@ with gr.Blocks() as demo:
185
  gr.Markdown("run Wan 2.2 in just 4-8 steps, with [Lightning LoRA](https://huggingface.co/Kijai/WanVideo_comfy/tree/main/Wan22-Lightning), fp8 quantization & AoT compilation - compatible with 🧨 diffusers and ZeroGPU⚡️")
186
  with gr.Row():
187
  with gr.Column():
188
- input_image_component = gr.Image(type="pil", label="Input Image (auto-resized to target H/W)")
189
  prompt_input = gr.Textbox(label="Prompt", value=default_prompt_i2v)
190
  duration_seconds_input = gr.Slider(minimum=MIN_DURATION, maximum=MAX_DURATION, step=0.1, value=3.5, label="Duration (seconds)", info=f"Clamped to model's {MIN_FRAMES_MODEL}-{MAX_FRAMES_MODEL} frames at {FIXED_FPS}fps.")
191
 
 
19
 
20
  MODEL_ID = "Wan-AI/Wan2.2-I2V-A14B-Diffusers"
21
 
22
+ MAX_DIM = 832
23
+ MIN_DIM = 480
24
+ SQUARE_DIM = 640
25
+ MULTIPLE_OF = 16
26
+
27
  MAX_SEED = np.iinfo(np.int32).max
28
 
29
  FIXED_FPS = 16
 
53
  torch.cuda.synchronize()
54
  torch.cuda.empty_cache()
55
 
56
+ OPTIMIZE_WIDTH = 832
57
+ OPTIMIZE_HEIGHT = 624
58
+
59
  optimize_pipeline_(pipe,
60
+ image=Image.new('RGB', (OPTIMIZE_WIDTH, OPTIMIZE_HEIGHT)),
61
  prompt='prompt',
62
+ height=OPTIMIZE_HEIGHT,
63
+ width=OPTIMIZE_WIDTH,
64
  num_frames=MAX_FRAMES_MODEL,
65
  )
66
 
 
68
  default_prompt_i2v = "make this image come alive, cinematic motion, smooth animation"
69
  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"
70
 
 
71
  def resize_image(image: Image.Image) -> Image.Image:
72
+ """
73
+ Resizes an image to fit within the model's constraints, preserving aspect ratio as much as possible.
74
+ """
75
+ width, height = image.size
 
76
 
77
+ # Handle square case
78
+ if width == height:
79
+ return image.resize((SQUARE_DIM, SQUARE_DIM), Image.LANCZOS)
80
 
81
+ aspect_ratio = width / height
82
+
83
+ MAX_ASPECT_RATIO = MAX_DIM / MIN_DIM
84
+ MIN_ASPECT_RATIO = MIN_DIM / MAX_DIM
85
+
86
+ image_to_resize = image
87
+
88
+ if aspect_ratio > MAX_ASPECT_RATIO:
89
+ # Very wide image -> crop width to fit 832x480 aspect ratio
90
+ target_w, target_h = MAX_DIM, MIN_DIM
91
+ crop_width = int(round(height * MAX_ASPECT_RATIO))
92
+ left = (width - crop_width) // 2
93
+ image_to_resize = image.crop((left, 0, left + crop_width, height))
94
+ elif aspect_ratio < MIN_ASPECT_RATIO:
95
+ # Very tall image -> crop height to fit 480x832 aspect ratio
96
+ target_w, target_h = MIN_DIM, MAX_DIM
97
+ crop_height = int(round(width / MIN_ASPECT_RATIO))
98
+ top = (height - crop_height) // 2
99
+ image_to_resize = image.crop((0, top, width, top + crop_height))
100
  else:
101
+ if width > height: # Landscape
102
+ target_w = MAX_DIM
103
+ target_h = int(round(target_w / aspect_ratio))
104
+ else: # Portrait
105
+ target_h = MAX_DIM
106
+ target_w = int(round(target_h * aspect_ratio))
107
+
108
+ final_w = round(target_w / MULTIPLE_OF) * MULTIPLE_OF
109
+ final_h = round(target_h / MULTIPLE_OF) * MULTIPLE_OF
110
+
111
+ final_w = max(MIN_DIM, min(MAX_DIM, final_w))
112
+ final_h = max(MIN_DIM, min(MAX_DIM, final_h))
113
+
114
+ return image_to_resize.resize((final_w, final_h), Image.LANCZOS)
115
+
116
 
117
  def get_duration(
118
  input_image,
 
176
  gr.Error: If input_image is None (no image uploaded).
177
 
178
  Note:
 
179
  - Frame count is calculated as duration_seconds * FIXED_FPS (24)
180
  - Output dimensions are adjusted to be multiples of MOD_VALUE (32)
181
  - The function uses GPU acceleration via the @spaces.GPU decorator
 
213
  gr.Markdown("run Wan 2.2 in just 4-8 steps, with [Lightning LoRA](https://huggingface.co/Kijai/WanVideo_comfy/tree/main/Wan22-Lightning), fp8 quantization & AoT compilation - compatible with 🧨 diffusers and ZeroGPU⚡️")
214
  with gr.Row():
215
  with gr.Column():
216
+ input_image_component = gr.Image(type="pil", label="Input Image")
217
  prompt_input = gr.Textbox(label="Prompt", value=default_prompt_i2v)
218
  duration_seconds_input = gr.Slider(minimum=MIN_DURATION, maximum=MAX_DURATION, step=0.1, value=3.5, label="Duration (seconds)", info=f"Clamped to model's {MIN_FRAMES_MODEL}-{MAX_FRAMES_MODEL} frames at {FIXED_FPS}fps.")
219
 
optimization.py CHANGED
@@ -14,18 +14,21 @@ from torchao.quantization import Int8WeightOnlyConfig
14
 
15
  from optimization_utils import capture_component_call
16
  from optimization_utils import aoti_compile
17
- from optimization_utils import ZeroGPUCompiledModel
18
  from optimization_utils import drain_module_parameters
19
 
20
 
21
  P = ParamSpec('P')
22
 
 
23
 
24
- TRANSFORMER_NUM_FRAMES_DIM = torch.export.Dim('num_frames', min=3, max=21)
 
25
 
26
  TRANSFORMER_DYNAMIC_SHAPES = {
27
  'hidden_states': {
28
- 2: TRANSFORMER_NUM_FRAMES_DIM,
 
 
29
  },
30
  }
31
 
@@ -44,6 +47,7 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
44
  @spaces.GPU(duration=1500)
45
  def compile_transformer():
46
 
 
47
  pipeline.load_lora_weights(
48
  "Kijai/WanVideo_comfy",
49
  weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors",
@@ -70,61 +74,33 @@ def optimize_pipeline_(pipeline: Callable[P, Any], *args: P.args, **kwargs: P.kw
70
  quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
71
  quantize_(pipeline.transformer_2, Float8DynamicActivationFloat8WeightConfig())
72
 
73
- hidden_states: torch.Tensor = call.kwargs['hidden_states']
74
- hidden_states_transposed = hidden_states.transpose(-1, -2).contiguous()
75
- if hidden_states.shape[-1] > hidden_states.shape[-2]:
76
- hidden_states_landscape = hidden_states
77
- hidden_states_portrait = hidden_states_transposed
78
- else:
79
- hidden_states_landscape = hidden_states_transposed
80
- hidden_states_portrait = hidden_states
81
-
82
- exported_landscape_1 = torch.export.export(
83
  mod=pipeline.transformer,
84
  args=call.args,
85
- kwargs=call.kwargs | {'hidden_states': hidden_states_landscape},
86
  dynamic_shapes=dynamic_shapes,
87
  )
88
 
89
- exported_portrait_2 = torch.export.export(
90
  mod=pipeline.transformer_2,
91
  args=call.args,
92
- kwargs=call.kwargs | {'hidden_states': hidden_states_portrait},
93
  dynamic_shapes=dynamic_shapes,
94
  )
95
 
96
- compiled_landscape_1 = aoti_compile(exported_landscape_1, INDUCTOR_CONFIGS)
97
- compiled_portrait_2 = aoti_compile(exported_portrait_2, INDUCTOR_CONFIGS)
 
 
98
 
99
- compiled_landscape_2 = ZeroGPUCompiledModel(compiled_landscape_1.archive_file, compiled_portrait_2.weights)
100
- compiled_portrait_1 = ZeroGPUCompiledModel(compiled_portrait_2.archive_file, compiled_landscape_1.weights)
101
-
102
- return (
103
- compiled_landscape_1,
104
- compiled_landscape_2,
105
- compiled_portrait_1,
106
- compiled_portrait_2,
107
- )
108
 
109
  quantize_(pipeline.text_encoder, Int8WeightOnlyConfig())
110
- cl1, cl2, cp1, cp2 = compile_transformer()
111
-
112
- def combined_transformer_1(*args, **kwargs):
113
- hidden_states: torch.Tensor = kwargs['hidden_states']
114
- if hidden_states.shape[-1] > hidden_states.shape[-2]:
115
- return cl1(*args, **kwargs)
116
- else:
117
- return cp1(*args, **kwargs)
118
-
119
- def combined_transformer_2(*args, **kwargs):
120
- hidden_states: torch.Tensor = kwargs['hidden_states']
121
- if hidden_states.shape[-1] > hidden_states.shape[-2]:
122
- return cl2(*args, **kwargs)
123
- else:
124
- return cp2(*args, **kwargs)
125
-
126
- pipeline.transformer.forward = combined_transformer_1
127
  drain_module_parameters(pipeline.transformer)
128
 
129
- pipeline.transformer_2.forward = combined_transformer_2
130
  drain_module_parameters(pipeline.transformer_2)
 
14
 
15
  from optimization_utils import capture_component_call
16
  from optimization_utils import aoti_compile
 
17
  from optimization_utils import drain_module_parameters
18
 
19
 
20
  P = ParamSpec('P')
21
 
22
+ LATENT_FRAMES_DIM = torch.export.Dim('num_latent_frames', min=8, max=81)
23
 
24
+ LATENT_PATCHED_HEIGHT_DIM = torch.export.Dim('latent_patched_height', min=30, max=52)
25
+ LATENT_PATCHED_WIDTH_DIM = torch.export.Dim('latent_patched_width', min=30, max=52)
26
 
27
  TRANSFORMER_DYNAMIC_SHAPES = {
28
  'hidden_states': {
29
+ 2: LATENT_FRAMES_DIM,
30
+ 3: 2 * LATENT_PATCHED_HEIGHT_DIM,
31
+ 4: 2 * LATENT_PATCHED_WIDTH_DIM,
32
  },
33
  }
34
 
 
47
  @spaces.GPU(duration=1500)
48
  def compile_transformer():
49
 
50
+ # This LoRA fusion part remains the same
51
  pipeline.load_lora_weights(
52
  "Kijai/WanVideo_comfy",
53
  weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors",
 
74
  quantize_(pipeline.transformer, Float8DynamicActivationFloat8WeightConfig())
75
  quantize_(pipeline.transformer_2, Float8DynamicActivationFloat8WeightConfig())
76
 
77
+
78
+ exported_1 = torch.export.export(
 
 
 
 
 
 
 
 
79
  mod=pipeline.transformer,
80
  args=call.args,
81
+ kwargs=call.kwargs,
82
  dynamic_shapes=dynamic_shapes,
83
  )
84
 
85
+ exported_2 = torch.export.export(
86
  mod=pipeline.transformer_2,
87
  args=call.args,
88
+ kwargs=call.kwargs,
89
  dynamic_shapes=dynamic_shapes,
90
  )
91
 
92
+ compiled_1 = aoti_compile(exported_1, INDUCTOR_CONFIGS)
93
+ compiled_2 = aoti_compile(exported_2, INDUCTOR_CONFIGS)
94
+
95
+ return compiled_1, compiled_2
96
 
 
 
 
 
 
 
 
 
 
97
 
98
  quantize_(pipeline.text_encoder, Int8WeightOnlyConfig())
99
+
100
+ compiled_transformer_1, compiled_transformer_2 = compile_transformer()
101
+
102
+ pipeline.transformer.forward = compiled_transformer_1
 
 
 
 
 
 
 
 
 
 
 
 
 
103
  drain_module_parameters(pipeline.transformer)
104
 
105
+ pipeline.transformer_2.forward = compiled_transformer_2
106
  drain_module_parameters(pipeline.transformer_2)