Mage-ViT is the visual encoder at the core of Mage-VL. It is a Codec-ViT built primarily for video, where a single image is simply the degenerate one-frame case. Mage-ViT follows a codec-aligned sparsity principle: visual tokens should be spent where a video codec spends bits, since codec bit-allocation is a natural proxy for spatio-temporal importance. On a shared 16×16 patch grid it keeps every anchor (I-frame) patch and only the motion-salient predicted (P-frame) patches, while a shared 3D rotary position encoding preserves spatio-temporal structure even after large fractions of the grid are dropped.
This repository is the ViT-pre-trained checkpoint only — it has not gone through the joint VLM training with the language model. Use it as a data-efficient, codec-native visual encoder, or as a drop-in ViT for your own multimodal training.
Codec-driven patchifier. Patches are selected by a per-patch importance map derived from the codec — motion vectors + P-frame residual energy for HEVC/H.265, or the learned rate map of the neural codec DCVC-RT. For a 64-frame clip it keeps all I-frame patches plus the top-k P-frame patches within a 4096-token budget (~75% token reduction). Chunk-wise and collage patchification are also supported.
Codec-agnostic. The same interface accepts a traditional codec (HEVC/H.265) or a neural codec (DCVC-RT) with no architecture or retraining change.
Trained from scratch. No billion-scale image-text ViT initialization — Mage-ViT is optimized with a large-scale cluster-discrimination objective on ~100M unlabeled images/videos.
A 24-layer pre-norm Vision Transformer trunk (hidden size 1024, 16 attention heads, GELU MLP at 4× expansion) processes the variable-length token sequence produced by the codec-driven patchifier, followed by a multi-head attention pooling head. The standalone encoder in this repo consumes pixel_values (and optional patch_positions); codec-driven patch selection is applied upstream in the Mage-VL data pipeline.
Pre-training is a two-stage, from-scratch recipe in bf16: (1) variable-resolution image pre-training (224–448), then (2) joint image + video pre-training (video at resolution 256, 64 frames per clip, 4096-token budget).
pip install "transformers>=5.7" torch torchvision pillow
# optional, for the fastest GPU attention path:
# pip install flash-attn --no-build-isolation
Attention is dispatched internally to sdpa (default) → flash_attention_2 → eager, so flash-attn is optional and the model runs on CPU or GPU out of the box.
Stack the preprocessed frames into [B, C, T, H, W] and pass a patch_positions tensor of shape [B, T * tokens_per_frame, 3] giving the (t, h, w) grid coordinate of every patch:
import torch
from PIL import Image
PATCH = 16
def build_patch_positions(num_frames, target_frames, grid_h, grid_w, device="cuda"):
# temporal index for each frame, spread across the target timeline
t = torch.linspace(0, target_frames - 1, num_frames, device=device).long()
t = t.repeat_interleave(grid_h * grid_w) # [T * H * W]
h = torch.arange(grid_h, device=device).repeat_interleave(grid_w).repeat(num_frames)
w = torch.arange(grid_w, device=device).repeat(grid_h).repeat(num_frames)
return torch.stack([t, h, w], dim=-1).unsqueeze(0) # [1, T*H*W, 3]
frames = [Image.open(f"frame_{i}.jpg") for i in range(16)] # your sampled frames
pv = processor(images=frames, return_tensors="pt")["pixel_values"] # [T, C, H, W]
video = pv.unsqueeze(0).permute(0, 2, 1, 3, 4).to("cuda") # [1, C, T, H, W]
gh, gw = video.shape[-2] // PATCH, video.shape[-1] // PATCH
patch_positions = build_patch_positions(num_frames=16, target_frames=64,
grid_h=gh, grid_w=gw)
with torch.no_grad():
out = model(video, patch_positions=patch_positions)
microsoft <huggingface.co/microsoft>Newsgroups: alt.freeweights.visionSubject: 
onnx-community