zerochocobo/matanyone2_onnx

MatAnyone2 ONNX

This repository contains fixed-shape ONNX exports of MatAnyone2 inference subgraphs.

The models are intended for projects that want to run MatAnyone2 with ONNX Runtime without importing PyTorch at runtime. They are especially useful for offline video matting, VR180/SBS workflows, and first-frame-mask propagation.

Upstream model: https://github.com/pq-yang/MatAnyone2

Model Folders

Each folder is self-contained and includes manifest.json.

matanyone2_onnx_512_bs1
matanyone2_onnx_512_bs2
matanyone2_onnx_1024_bs1
matanyone2_onnx_1024_bs2

Use 512 for faster processing and lower VRAM. Use 1024 for higher-resolution masks when your GPU can handle the extra cost.

Use bs1 for normal single-view video or sequential left/right-eye processing.

Use bs2 when you want to process side-by-side VR180 left/right eyes as a batch of 2.

Files

Required files:

manifest.json
matanyone2_image_key.onnx
matanyone2_mask_memory.onnx
matanyone2_first_frame_refine.onnx
matanyone2_propagate.onnx

Optional acceleration/convenience files:

matanyone2_propagate_update.onnx
matanyone2_step_update.onnx

step_update fuses image feature extraction, propagation, and memory update for following frames. If it is missing, callers can run image_key, propagate, and mask_memory separately.

Input Preprocessing

Video frame input to matanyone2_image_key.onnx:

shape: [batch, 3, height, width]
layout: RGB, CHW
dtype: float32 unless the ONNX file input says float16
range: 0.0 to 1.0

First-frame mask input to matanyone2_mask_memory.onnx:

shape: [batch, 1, height, width]
dtype: same as image input
range: 0.0 background, 1.0 foreground

The exported height and width are fixed. Read them from manifest.json.

Runtime Dependencies

GPU:

pip install onnxruntime-gpu opencv-python numpy

CPU smoke tests:

pip install onnxruntime opencv-python numpy

Standalone Video Example

This repository includes matanyone2_onnx_video_infer.py, a complete example that loads all exported ONNX graphs, propagates a first-frame mask through a video, and writes a preview video.

Single-view alpha-mask output:

python matanyone2_onnx_video_infer.py \
  --model-dir matanyone2_onnx_512_bs1 \
  --video input.mp4 \
  --mask first_frame_mask.png \
  --output-mode alpha \
  --out alpha_preview.mp4

Side-by-side VR180 green-background preview:

python matanyone2_onnx_video_infer.py \
  --model-dir matanyone2_onnx_512_bs2 \
  --video sbs_vr180.mp4 \
  --mask first_frame_mask_sbs.png \
  --sbs \
  --output-mode green \
  --out green_preview.mp4

Use a 1024 model:

python matanyone2_onnx_video_infer.py \
  --model-dir matanyone2_onnx_1024_bs1 \
  --video input.mp4 \
  --mask first_frame_mask.png \
  --out alpha_1024.mp4

ONNX Graph Call Order

For the first frame:

image_key(image)
mask_memory(image, first_frame_mask, zero_sensory, pix_feat)
first_frame_refine(features, msk_value, obj_memory, sensory, first_frame_mask)
mask_memory(image, refined_alpha, sensory, pix_feat)

For following frames, prefer:

step_update(image, memory_key, memory_shrinkage, msk_value, obj_memory, sensory, last_mask, last_pix_feat, last_pred_mask, last_msk_value)

Fallback when step_update is unavailable:

image_key(image)
propagate(features, memory_key, memory_shrinkage, msk_value, obj_memory, sensory, last_mask, last_pix_feat, last_pred_mask, last_msk_value)
mask_memory(image, alpha, sensory, pix_feat)

The example script implements both paths.

Notes

• The current exports target one foreground object.
• The models do not generate the first mask. You must provide a first-frame mask from SAM, manual annotation, chroma keying, or another detector.
• For long videos or scene changes, reset the state periodically and provide a new bootstrap mask.
• bs2 is useful for left/right eye batching, but it also increases memory pressure.
• These ONNX files are exported from the upstream MatAnyone2 model and should be used under the upstream model's license and terms.