Image-Text-to-Text
Transformers
Safetensors
mage_vl
multimodal
vision-language-model
mage-vl
video-understanding
streaming
conversational
custom_code
Instructions to use microsoft/Mage-VL with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use microsoft/Mage-VL with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-text-to-text", model="microsoft/Mage-VL", trust_remote_code=True) messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] pipe(text=messages)# Load model directly from transformers import AutoModelForImageTextToText model = AutoModelForImageTextToText.from_pretrained("microsoft/Mage-VL", trust_remote_code=True, device_map="auto") - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use microsoft/Mage-VL with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "microsoft/Mage-VL" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "microsoft/Mage-VL", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker
docker model run hf.co/microsoft/Mage-VL
- SGLang
How to use microsoft/Mage-VL with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "microsoft/Mage-VL" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "microsoft/Mage-VL", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "microsoft/Mage-VL" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "microsoft/Mage-VL", "messages": [ { "role": "user", "content": [ { "type": "text", "text": "Describe this image in one sentence." }, { "type": "image_url", "image_url": { "url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg" } } ] } ] }' - Docker Model Runner
How to use microsoft/Mage-VL with Docker Model Runner:
docker model run hf.co/microsoft/Mage-VL
File size: 17,735 Bytes
12acbba | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 | #!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""Frame processing utilities including resize, letterbox detection, and bad frame filtering."""
from typing import Tuple, Optional, List
import subprocess
import numpy as np
import cv2
def _resize_bgr(frame_bgr: np.ndarray, out_h: int, out_w: int) -> np.ndarray:
"""Resize BGR frame with area when shrinking and linear when enlarging."""
H, W = frame_bgr.shape[:2]
if int(H) == int(out_h) and int(W) == int(out_w):
return frame_bgr
shrink = (out_h < H) or (out_w < W)
interp = cv2.INTER_AREA if shrink else cv2.INTER_LINEAR
return cv2.resize(frame_bgr, (int(out_w), int(out_h)), interpolation=interp)
def _resize_gray(gray: np.ndarray, out_h: int, out_w: int) -> np.ndarray:
"""Resize single-channel image with area when shrinking and linear when enlarging."""
H, W = gray.shape[:2]
if int(H) == int(out_h) and int(W) == int(out_w):
return gray
shrink = (out_h < H) or (out_w < W)
interp = cv2.INTER_AREA if shrink else cv2.INTER_LINEAR
return cv2.resize(gray, (int(out_w), int(out_h)), interpolation=interp)
def _resize_mv_and_scale(mv: np.ndarray, out_h: int, out_w: int, scale_h: float, scale_w: float) -> np.ndarray:
"""Resize MV field to (out_h,out_w) and scale vector components.
Assumption (as per user): mv/res are pixel-aligned. When spatially resizing the
image, motion in pixel units should be scaled by the same factors.
Supports mv shape (...,2) or (...,4) where channels are (x,y) or (l0x,l0y,l1x,l1y).
"""
if mv.ndim != 3 or mv.shape[2] < 2:
return mv
Hm, Wm = mv.shape[:2]
if int(Hm) == int(out_h) and int(Wm) == int(out_w):
mv_rs = mv.astype(np.float32, copy=False)
else:
# resize per-channel
mv_f = mv.astype(np.float32, copy=False)
ch = mv_f.shape[2]
mv_rs = np.zeros((int(out_h), int(out_w), ch), dtype=np.float32)
for c in range(ch):
mv_rs[:, :, c] = cv2.resize(mv_f[:, :, c], (int(out_w), int(out_h)), interpolation=cv2.INTER_LINEAR)
# scale vector components
mv_rs[:, :, 0] *= float(scale_w)
mv_rs[:, :, 1] *= float(scale_h)
if mv_rs.shape[2] >= 4:
mv_rs[:, :, 2] *= float(scale_w)
mv_rs[:, :, 3] *= float(scale_h)
return mv_rs
def detect_letterbox_bbox_bgr(frame_bgr: np.ndarray, dark_thr: float = 16.0) -> Tuple[int, int, int, int]:
"""Detect content bbox [top,bottom,left,right] by scanning dark borders."""
if frame_bgr is None or frame_bgr.size == 0:
return 0, 0, 0, 0
h, w = frame_bgr.shape[:2]
y = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2YUV)[:, :, 0].astype(np.float32)
row_mean = y.mean(axis=1)
col_mean = y.mean(axis=0)
max_tb = int(h * 0.45)
max_lr = int(w * 0.45)
top = 0
while top < min(h, max_tb) and float(row_mean[top]) < float(dark_thr):
top += 1
bottom = h
while bottom - 1 >= max(0, h - max_tb) and float(row_mean[bottom - 1]) < float(dark_thr):
bottom -= 1
left = 0
while left < min(w, max_lr) and float(col_mean[left]) < float(dark_thr):
left += 1
right = w
while right - 1 >= max(0, w - max_lr) and float(col_mean[right - 1]) < float(dark_thr):
right -= 1
if top >= bottom or left >= right:
return 0, h, 0, w
return int(top), int(bottom), int(left), int(right)
def _bgr_to_luma_u8(frame_bgr: np.ndarray) -> np.ndarray:
"""Convert BGR uint8 frame to luma (Y) uint8."""
if frame_bgr is None or frame_bgr.size == 0:
return np.zeros((0, 0), dtype=np.uint8)
if frame_bgr.ndim == 2:
return frame_bgr.astype(np.uint8, copy=False)
if frame_bgr.shape[2] == 1:
return frame_bgr[:, :, 0].astype(np.uint8, copy=False)
yuv = cv2.cvtColor(frame_bgr, cv2.COLOR_BGR2YUV)
return yuv[:, :, 0].astype(np.uint8, copy=False)
def is_black_frame(
frame_bgr: np.ndarray,
y_mean_thr: float = 8.0,
y_std_thr: float = 6.0,
) -> bool:
"""Detect near-black (contentless) frames.
Heuristic:
- mean(Y) very low AND std(Y) very low.
This filters pure black fades/blank segments.
"""
try:
y = _bgr_to_luma_u8(frame_bgr)
if y.size == 0:
return True
m = float(y.mean())
s = float(y.std())
return (m <= float(y_mean_thr)) and (s <= float(y_std_thr))
except Exception:
return False
def is_solid_color_frame(
frame_bgr: np.ndarray,
y_std_thr: float = 4.0,
color_std_thr: float = 4.0,
max_range_thr: float = 12.0,
) -> bool:
"""Detect nearly pure-color / flat frames.
This is broader than black-frame detection and catches frames that are almost
entirely one color, such as pure white / gray / green screens or flat fades.
Heuristic:
- luma std is very low
- per-channel std is very low
- dynamic range (max-min) of each channel is small
"""
try:
if frame_bgr is None or frame_bgr.size == 0:
return True
if frame_bgr.ndim != 3 or frame_bgr.shape[2] < 3:
y = _bgr_to_luma_u8(frame_bgr)
if y.size == 0:
return True
return bool((float(y.std()) <= float(y_std_thr)) and ((float(y.max()) - float(y.min())) <= float(max_range_thr)))
y = _bgr_to_luma_u8(frame_bgr)
if y.size == 0:
return True
b = frame_bgr[:, :, 0].astype(np.float32, copy=False)
g = frame_bgr[:, :, 1].astype(np.float32, copy=False)
r = frame_bgr[:, :, 2].astype(np.float32, copy=False)
y_std = float(y.std())
b_std = float(b.std())
g_std = float(g.std())
r_std = float(r.std())
b_rng = float(b.max() - b.min())
g_rng = float(g.max() - g.min())
r_rng = float(r.max() - r.min())
return bool(
(y_std <= float(y_std_thr))
and (b_std <= float(color_std_thr))
and (g_std <= float(color_std_thr))
and (r_std <= float(color_std_thr))
and (b_rng <= float(max_range_thr))
and (g_rng <= float(max_range_thr))
and (r_rng <= float(max_range_thr))
)
except Exception:
return False
def is_corrupted_green_frame(
frame_bgr: np.ndarray,
green_frac_thr: float = 0.35,
g_thr: int = 180,
rb_thr: int = 90,
) -> bool:
"""Detect typical FFmpeg/OpenCV decode corruption that manifests as large green blocks.
Heuristic:
- Count pixels with very strong G while R/B are low.
- If the fraction is high, treat frame as corrupted.
This catches the common 'green macroblock' artifacts (like the screenshot).
"""
try:
if frame_bgr is None or frame_bgr.size == 0:
return True
if frame_bgr.ndim != 3 or frame_bgr.shape[2] < 3:
return False
b = frame_bgr[:, :, 0].astype(np.uint8, copy=False)
g = frame_bgr[:, :, 1].astype(np.uint8, copy=False)
r = frame_bgr[:, :, 2].astype(np.uint8, copy=False)
mask = (g >= int(g_thr)) & (r <= int(rb_thr)) & (b <= int(rb_thr))
frac = float(mask.mean())
if frac >= float(green_frac_thr):
return True
# Secondary check: overall green dominance with low texture (often corrupted fill)
mg = float(g.mean()); mr = float(r.mean()); mb = float(b.mean())
sg = float(g.std())
if (mg > 120.0) and (mg - max(mr, mb) > 80.0) and (sg < 50.0):
return True
return False
except Exception:
return False
def frame_is_bad(
frame_bgr: np.ndarray,
skip_black_frames: bool = True,
skip_corrupt_frames: bool = True,
black_y_mean_thr: float = 8.0,
black_y_std_thr: float = 6.0,
solid_y_std_thr: float = 4.0,
solid_color_std_thr: float = 4.0,
solid_max_range_thr: float = 12.0,
corrupt_green_frac_thr: float = 0.35,
corrupt_g_thr: int = 180,
corrupt_rb_thr: int = 90,
) -> bool:
"""Unified bad-frame predicate used to avoid selecting blank/corrupted segments."""
if skip_black_frames and is_black_frame(frame_bgr, y_mean_thr=black_y_mean_thr, y_std_thr=black_y_std_thr):
return True
if is_solid_color_frame(
frame_bgr,
y_std_thr=solid_y_std_thr,
color_std_thr=solid_color_std_thr,
max_range_thr=solid_max_range_thr,
):
return True
if skip_corrupt_frames and is_corrupted_green_frame(
frame_bgr,
green_frac_thr=corrupt_green_frac_thr,
g_thr=corrupt_g_thr,
rb_thr=corrupt_rb_thr,
):
return True
return False
def pad_to_multiple_of_bgr(frame_bgr: np.ndarray, base: int) -> Tuple[np.ndarray, Tuple[int, int]]:
"""Pad bottom/right with zeros so that H and W are multiples of `base`.
For patch-based 2x2 blocks, you typically want base = 2 * patch.
Returns padded frame and (pad_bottom, pad_right).
"""
base = int(max(1, base))
H, W = frame_bgr.shape[:2]
pad_bottom = (base - (H % base)) % base
pad_right = (base - (W % base)) % base
if pad_bottom == 0 and pad_right == 0:
return frame_bgr, (0, 0)
out = cv2.copyMakeBorder(
frame_bgr,
0,
pad_bottom,
0,
pad_right,
borderType=cv2.BORDER_CONSTANT,
value=(0, 0, 0),
)
return out, (pad_bottom, pad_right)
def pad_to_multiple_of_32_bgr(frame_bgr: np.ndarray) -> Tuple[np.ndarray, Tuple[int, int]]:
"""Pad BGR frame so both dimensions are multiples of 32."""
return pad_to_multiple_of_bgr(frame_bgr, 32)
def bgr_to_residual_y_u8(frame_bgr: np.ndarray) -> np.ndarray:
"""Fallback residual proxy when cv_reader residual is unavailable.
IMPORTANT:
- `mv_res_score_map()` expects a residual-like signal centered at 128, and uses
`abs(res_y - 128)` as energy.
- Using raw luma Y directly will incorrectly give high energy to pure black/white
regions (|Y-128| is large), causing many black patches to be selected.
This proxy uses Sobel gradient magnitude on luma as a texture/edge energy estimate,
then maps it into uint8 residual space with 128 as the zero point:
res_proxy = 128 + clip(grad_norm * 127)
Flat regions (black/white) have low gradient, so they no longer dominate.
"""
if frame_bgr is None or frame_bgr.size == 0:
return np.full((0, 0), 128, dtype=np.uint8)
# Luma
y = _bgr_to_luma_u8(frame_bgr).astype(np.float32)
if y.size == 0:
return np.full((0, 0), 128, dtype=np.uint8)
# Sobel gradients (float32)
gx = cv2.Sobel(y, cv2.CV_32F, 1, 0, ksize=3)
gy = cv2.Sobel(y, cv2.CV_32F, 0, 1, ksize=3)
mag = cv2.magnitude(gx, gy) # >=0
# Robust normalization to [0,1] using percentile to avoid outliers
scale = float(np.percentile(mag, 95.0))
if (not np.isfinite(scale)) or scale <= 1e-6:
scale = 1.0
mag_n = np.clip(mag / scale, 0.0, 1.0)
# Map to residual-like uint8 with 128 as center
res_proxy = (128.0 + mag_n * 127.0).astype(np.uint8)
return res_proxy
def decode_frame_bgr_at(video_path: str, frame_id: int, backsearch_max: int = 32) -> Optional[np.ndarray]:
"""Decode a single frame at given frame_id.
If the exact frame fails, try nearby frames up to backsearch_max.
Returns:
BGR frame or None if decoding fails.
"""
cap = cv2.VideoCapture(video_path)
if not cap.isOpened():
return None
# Try exact frame first
cap.set(cv2.CAP_PROP_POS_FRAMES, frame_id)
ret, frame = cap.read()
if ret and frame is not None and frame.size > 0:
cap.release()
return frame
# Try nearby frames
for offset in range(1, backsearch_max + 1):
for sign in [1, -1]:
try_id = frame_id + sign * offset
if try_id < 0:
continue
cap.set(cv2.CAP_PROP_POS_FRAMES, try_id)
ret, frame = cap.read()
if ret and frame is not None and frame.size > 0:
cap.release()
return frame
cap.release()
return None
def decode_frames_bgr_opencv(video_path: str, frame_ids: List[int], backsearch_max: int = 32) -> List[np.ndarray]:
"""Decode multiple frames using a single VideoCapture (faster).
Important:
- When a stream is partially corrupt, OpenCV/FFmpeg may fail at some frame ids.
Doing an unbounded decrement loop can become O(video_length) per failed frame
and destroy throughput. We cap the backsearch and otherwise reuse last_good.
"""
cap = cv2.VideoCapture(str(video_path))
if not cap.isOpened():
raise RuntimeError(f"Cannot open video: {video_path}")
total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
out: List[np.ndarray] = []
last_good: Optional[np.ndarray] = None
bs = int(max(0, backsearch_max))
for fid0 in frame_ids:
fid = int(fid0)
if total > 0:
fid = max(0, min(fid, total - 1))
cap.set(cv2.CAP_PROP_POS_FRAMES, fid)
ok, frame = cap.read()
if (not ok) or frame is None:
# bounded backward search
if bs > 0:
dec = fid
tries = 0
frame = None
while dec > 0 and tries < bs:
dec -= 1
tries += 1
cap.set(cv2.CAP_PROP_POS_FRAMES, dec)
ok2, f2 = cap.read()
if ok2 and f2 is not None:
frame = f2
break
# if still none, reuse last_good immediately (fast)
if frame is None:
if last_good is None:
cap.release()
raise RuntimeError(f"Failed to decode any frame around fid={fid0} for {video_path}")
frame = last_good
else:
last_good = frame
out.append(frame)
cap.release()
return out
def decode_frames_bgr_ffmpeg_subprocess(video_path: str, frame_ids: List[int]) -> List[np.ndarray]:
"""Decode selected frames using system ffmpeg subprocess and rawvideo pipe."""
frame_ids = [int(x) for x in frame_ids]
if not frame_ids:
return []
cap = cv2.VideoCapture(str(video_path))
if not cap.isOpened():
raise RuntimeError(f"Cannot open video metadata: {video_path}")
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
total = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
cap.release()
if width <= 0 or height <= 0:
raise RuntimeError(f"Invalid video size for {video_path}: {width}x{height}")
uniq_ids = sorted(set(max(0, min(int(fid), total - 1)) if total > 0 else max(0, int(fid))
for fid in frame_ids))
if not uniq_ids:
return []
select_expr = "+".join(f"eq(n\\,{fid})" for fid in uniq_ids)
cmd = [
"ffmpeg",
"-v", "error",
"-i", str(video_path),
"-vf", f"select={select_expr}",
"-vsync", "0",
"-f", "rawvideo",
"-pix_fmt", "bgr24",
"-",
]
proc = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
if proc.returncode != 0:
err = proc.stderr.decode("utf-8", errors="ignore")
raise RuntimeError(f"ffmpeg decode failed: {err.strip()}")
frame_size = int(width) * int(height) * 3
if frame_size <= 0:
raise RuntimeError(f"Invalid frame_size for {video_path}")
raw = proc.stdout
num_decoded = len(raw) // frame_size
if num_decoded <= 0:
raise RuntimeError(f"ffmpeg returned no frames for {video_path}")
if len(raw) % frame_size != 0:
raw = raw[: num_decoded * frame_size]
decoded_map = {}
for i in range(min(num_decoded, len(uniq_ids))):
start = i * frame_size
end = start + frame_size
arr = np.frombuffer(raw[start:end], dtype=np.uint8).copy().reshape(height, width, 3)
decoded_map[int(uniq_ids[i])] = arr
out: List[np.ndarray] = []
last_good: Optional[np.ndarray] = None
for fid in frame_ids:
key = max(0, min(int(fid), total - 1)) if total > 0 else max(0, int(fid))
fr = decoded_map.get(key)
if fr is None:
if last_good is None:
# fallback: pick nearest decoded frame
nearest = None
if decoded_map:
nearest = decoded_map[min(decoded_map.keys(), key=lambda x: abs(x - key))]
if nearest is None:
raise RuntimeError(f"missing decoded frame for fid={fid} in {video_path}")
fr = nearest
else:
fr = last_good
out.append(np.ascontiguousarray(fr))
last_good = fr
return out
def decode_frames_bgr(
video_path: str,
frame_ids: List[int],
backsearch_max: int = 32,
backend: str = "auto",
) -> List[np.ndarray]:
"""Decode multiple frames with configurable backend.
backend:
- auto: prefer system ffmpeg subprocess, fallback to OpenCV
- ffmpeg_native: use system ffmpeg subprocess and raise on failure
- opencv: force OpenCV VideoCapture path
"""
backend = str(backend).lower().strip()
if backend not in {"auto", "ffmpeg_native", "opencv"}:
backend = "auto"
if backend in {"auto", "ffmpeg_native"}:
try:
return decode_frames_bgr_ffmpeg_subprocess(video_path, frame_ids)
except Exception:
if backend == "ffmpeg_native":
raise
return decode_frames_bgr_opencv(video_path, frame_ids, backsearch_max=backsearch_max)
|