File size: 11,403 Bytes
8603681
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
26b94a3
8603681
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
import importlib
import os
import subprocess
import sys
import tempfile
from pathlib import Path

# Install videoflextok without its deps to avoid huggingface_hub==0.25.2 conflicting
# with gradio's >=0.33.5 requirement. Compatible dep versions are in requirements.txt.
def _install_videoflextok():
    try:
        import videoflextok  # noqa: F401
        return
    except ImportError:
        pass
    print("[VideoFlexTok] Installing videoflextok (--no-deps) ...")
    subprocess.run(
        [sys.executable, "-m", "pip", "install", "--quiet", "--no-deps",
         "git+https://github.com/apple/ml-videoflextok.git"],
        check=True,
    )
    importlib.invalidate_caches()

_install_videoflextok()

import spaces
import gradio as gr
import imageio.v3 as iio
import numpy as np
import torch

from videoflextok.utils.demo import denormalize, read_mp4
from videoflextok.utils.misc import detect_bf16_support, get_bf16_context
from videoflextok.wrappers import VideoFlexTokFromHub


# --- Constants ---------------------------------------------------------------------

MODEL_ID = "EPFL-VILAB/videoflextok_d18_d28"
APP_DIR = Path(__file__).resolve().parent
EXAMPLES_DIR = APP_DIR / "examples"
EXAMPLE_VIDEOS = sorted(EXAMPLES_DIR.glob("*.mp4"))
NUM_KEEP_TOKENS = [2**i for i in range(9)]  # 1, 2, 4, 8, 16, 32, 64, 128, 256

APP_CSS = """
#col-container {
  margin: 0 auto;
  max-width: 1500px;
}
#col-input-container {
  margin: 0 auto;
  max-width: 420px;
}
#run-button {
  margin: 0 auto;
}
"""


# --- Device setup ------------------------------------------------------------------

torch.set_grad_enabled(False)
if torch.cuda.is_available():
    torch.backends.cuda.matmul.allow_tf32 = True
    torch.backends.cudnn.allow_tf32 = True

DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
ENABLE_BF16 = DEVICE.type == "cuda" and detect_bf16_support()


# --- Model loading -----------------------------------------------------------------

def _patch_for_hf_spaces(model):
    """Patch TorchDynamo and model for HF Spaces / ZeroGPU compatibility.

    This PyTorch version's TorchDynamo cannot represent torch.device as a ConstantVariable,
    causing torch.compile(flex_attention) to crash. The fix was merged into newer PyTorch;
    here we backport it by adding torch.device to common_constant_types, so the Triton
    kernel is used correctly instead of falling back to the dense O(n²) math implementation.

    We also disable block mask compilation (compile_block_mask=False) since create_block_mask
    uses a separate internal torch.compile call that would hit the same bug.
    """
    # Patch TorchDynamo to accept torch.device as a ConstantVariable.
    # common_constant_types may be closed over in is_base_literal, so patch the method directly.
    import torch._dynamo.variables.constant as _dynamo_const
    _orig_is_base_literal = _dynamo_const.ConstantVariable.is_base_literal

    @staticmethod
    def _patched_is_base_literal(value):
        return isinstance(value, torch.device) or _orig_is_base_literal(value)

    _dynamo_const.ConstantVariable.is_base_literal = _patched_is_base_literal

    from videoflextok.model.preprocessors.flex_seq_packing import (
        BlockWiseSequencePacker,
        BlockWiseSequenceInterleavePacker,
        BlockWiseSequencePackerWithCrossAttention,
    )
    for module in model.modules():
        if isinstance(module, (
            BlockWiseSequencePacker,
            BlockWiseSequenceInterleavePacker,
            BlockWiseSequencePackerWithCrossAttention,
        )):
            module.compile_block_mask = False


_model = None
try:
    print(f"[VideoFlexTok] Loading {MODEL_ID} ...")
    _model = VideoFlexTokFromHub.from_pretrained(MODEL_ID)
    _model = _model.to(torch.bfloat16).to(DEVICE).eval()
    _patch_for_hf_spaces(_model)
    print("[VideoFlexTok] Model ready.")
except Exception as exc:
    print(f"[VideoFlexTok] FATAL: model load failed: {exc}")


# --- Inference ---------------------------------------------------------------------

def _stack_reconstructed_videos(videos, output_path: str, fps: int):
    """Compose 9 reconstructions + original into a 2×5 grid video and write to output_path."""
    def to_uint8_frames(video_tensor):
        if video_tensor.ndim == 5:
            video_tensor = video_tensor[0]
        frames = denormalize(video_tensor).permute(1, 2, 3, 0).contiguous().numpy()
        return (np.clip(frames, 0.0, 1.0) * 255).round().astype(np.uint8)

    def add_border(frames: np.ndarray, border_px: int, color: int) -> np.ndarray:
        return np.pad(
            frames,
            ((0, 0), (border_px, border_px), (border_px, border_px), (0, 0)),
            mode="constant", constant_values=color,
        )

    def compose_row(row_frames: list[np.ndarray], t: int, gap_px: int) -> np.ndarray:
        gap_col = np.full((row_frames[0].shape[1], gap_px, 3), 255, dtype=np.uint8)
        items = []
        for i, frames in enumerate(row_frames):
            items.append(frames[t])
            if i < len(row_frames) - 1:
                items.append(gap_col)
        return np.concatenate(items, axis=1)

    border_px, gap_px = 8, 8
    reconstructed = [add_border(to_uint8_frames(v), border_px, 255) for v in videos[:9]]
    original = add_border(to_uint8_frames(videos[9]), border_px, 0)

    all_panels = reconstructed + [original]
    total_frames = min(p.shape[0] for p in all_panels)
    all_panels = [p[:total_frames] for p in all_panels]

    row1 = all_panels[:5]   # k = 1, 2, 4, 8, 16
    row2 = all_panels[5:]   # k = 32, 64, 128, 256, Original

    composed = []
    for t in range(total_frames):
        row1_img = compose_row(row1, t, gap_px)
        row2_img = compose_row(row2, t, gap_px)
        row_gap = np.full((gap_px, row1_img.shape[1], 3), 255, dtype=np.uint8)
        composed.append(np.concatenate([row1_img, row_gap, row2_img], axis=0))

    iio.imwrite(
        output_path, np.stack(composed, axis=0),
        fps=fps, plugin="FFMPEG", codec="libx264", pixelformat="yuv420p",
    )


def reconstruct_video(video_path: str, input_fps: int, timesteps: int, guidance_scale: float, seed: int):
    if not video_path or not Path(video_path).exists():
        raise gr.Error("Upload a video first.")
    if _model is None:
        raise gr.Error("Model failed to load at startup — check Space logs.")

    try:
        preprocess_args = dict(_model.video_preprocess_args)
        # Public package uses 'overlap_size'; model config key is 'overlap_size_frames'
        if "overlap_size_frames" in preprocess_args and "overlap_size" not in preprocess_args:
            preprocess_args["overlap_size"] = preprocess_args.pop("overlap_size_frames")
        video_tensor = read_mp4(str(video_path), fps=int(input_fps), **preprocess_args)
    except Exception as exc:
        raise gr.Error(f"Failed to decode video: {exc}") from exc

    try:
        with get_bf16_context(ENABLE_BF16, device_type=DEVICE.type):
            print(f"[VideoFlexTok] Tokenizing {video_tensor.shape} ...")
            token_ids = _model.tokenize(video_tensor[None].to(DEVICE))
            print(f"[VideoFlexTok] Decoding {len(NUM_KEEP_TOKENS)} reconstructions ...")
            reconstructed = _model.detokenize(
                [token_ids[0]] * len(NUM_KEEP_TOKENS),
                num_keep_tokens_list=NUM_KEEP_TOKENS,
                timesteps=int(timesteps),
                guidance_scale=float(guidance_scale),
                perform_norm_guidance=True,
                generator=torch.Generator(device=DEVICE.type).manual_seed(int(seed)),
                eta=0.0, momentum=0.0, norm_threshold=0.6, verbose=False,
            )
        reconstructed = [v.cpu().float() for v in reconstructed]
        print("[VideoFlexTok] Inference complete.")
    except Exception as exc:
        raise gr.Error(f"Model inference failed: {exc}") from exc

    tmp = tempfile.NamedTemporaryFile(suffix=".mp4", delete=False)
    tmp.close()
    _stack_reconstructed_videos(reconstructed + [video_tensor], output_path=tmp.name, fps=int(input_fps))

    info = f"Extracted {video_tensor.shape[1]} frames at {input_fps} FPS"
    return tmp.name, info


if spaces is not None and hasattr(spaces, "GPU"):
    reconstruct_video = spaces.GPU(duration=60)(reconstruct_video)


# --- UI ----------------------------------------------------------------------------

with gr.Blocks(title="VideoFlexTok Demo", theme=gr.themes.Base(), css=APP_CSS) as demo:
    with gr.Column(elem_id="col-container"):
        gr.Markdown("# VideoFlexTok: Flexible-Length Coarse-to-Fine Video Tokenization")

        with gr.Row():
            with gr.Column(scale=1, elem_id="col-input-container"):
                gr.Markdown(f"""
[`Website`](https://videoflextok.epfl.ch) | [`Paper`](https://arxiv.org/abs/2604.12887) | [`GitHub`](https://github.com/apple/ml-videoflextok) | [`Model`](https://huggingface.co/EPFL-VILAB/videoflextok_d18_d28)

Research demo for **VideoFlexTok: Flexible-Length Coarse-to-Fine Video Tokenization** (arXiv 2026).
Autoencodes your video with `{MODEL_ID}` and shows coarse-to-fine reconstructions.
VideoFlexTok tokenizes video into `T × 256` tokens ordered coarse-to-fine; this demo shows
reconstructions from `T × k` tokens for k ∈ `{NUM_KEEP_TOKENS}`. Bottom-right is the original.
""")
                input_video = gr.Video(
                    label="Input video", sources=["upload"], format="mp4",
                )
                run_button = gr.Button("Autoencode with VideoFlexTok", elem_id="run-button")

                if EXAMPLE_VIDEOS:
                    gr.Examples(
                        examples=[str(p) for p in EXAMPLE_VIDEOS],
                        inputs=[input_video],
                        outputs=[input_video],
                        fn=lambda p: p,
                        cache_examples=True,
                        label="Example videos",
                    )

                with gr.Accordion("Advanced Settings", open=False):
                    gr.Markdown("Adjust target FPS to control how many frames are extracted.")
                    input_fps = gr.Slider(minimum=1, maximum=16, value=8, step=1, label="Target FPS")
                    timesteps = gr.Slider(minimum=1, maximum=60, value=20, step=1, label="Denoising steps")
                    guidance_scale = gr.Slider(minimum=1.0, maximum=30.0, value=25.0, step=0.5, label="Guidance scale")
                    seed = gr.Number(value=42, precision=0, label="Seed")

            with gr.Column(scale=4):
                output_video = gr.Video(label="Reconstructions")
                status = gr.Markdown()

    run_button.click(
        fn=reconstruct_video,
        inputs=[input_video, input_fps, timesteps, guidance_scale, seed],
        outputs=[output_video, status],
    )

    if DEVICE.type != "cuda":
        gr.Markdown("Running on CPU — inference will be slow.")


# --- Launch ------------------------------------------------------------------------

demo.queue(max_size=16)

if __name__ == "__main__":
    server_name = os.environ.get("GRADIO_SERVER_NAME", "0.0.0.0")
    launch_kwargs = {"server_name": server_name, "ssr_mode": False}
    if port := os.environ.get("GRADIO_SERVER_PORT"):
        launch_kwargs["server_port"] = int(port)
    launch_kwargs["allowed_paths"] = [str(APP_DIR), tempfile.gettempdir()]
    demo.launch(**launch_kwargs)