tribev2-high-quality-qv

This repo packages the full TRIBEv2 model with a quantized V-JEPA2 video branch.

The original TRIBEv2 checkpoint is still there, and the quantized video weights are already included in quantized_video/, so the repo stays easy to move around and use as a single package. The bundled loader keeps the original V-JEPA2 backbone id visible to TRIBEv2 consumers and swaps in the embedded quantized weights from quantized_video/ at load time. All presets in this family keep a ViT-G-compatible video feature interface so the untouched TRIBEv2 checkpoint remains dimensionally compatible.

This version

Highest-sampling runtime profile, using the same ViT-G-compatible quantized branch that matches the original checkpoint dimensions.

Good fit for:

• High-end GPUs
• Best-fidelity sharing
• Reference-like public variant

In short:

• Quality: Highest expected fidelity of the 3 presets.
• Memory: Heaviest runtime profile of the family.
• Ease of use: Usually the least convenient to run of the 3 presets.

Runtime profile

• Profile: High-quality profile near the original TRIBEv2 video sampling defaults.
• Video backbone: facebook/vjepa2-vitg-fpc64-256
• Video event frequency: 2.0 Hz
• Video batch size: 8
• Video num_frames: default

What you get

• Full TRIBEv2 repo in one place
• Original best.ckpt
• Audio, text and fusion parts left as-is
• Quantized video branch in quantized_video/
• Loader helper in load_quantized_tribev2.py that injects the embedded quantized video branch

Quick start

pip install "torch>=2.11" transformers huggingface_hub pyyaml torchao tribev2

Use the bundled loader below rather than calling TribeModel.from_pretrained(...) directly on the repo.

from pathlib import Path
from load_quantized_tribev2 import load_quantized_tribev2

repo_dir = Path(".")
model = load_quantized_tribev2(repo_dir, device="auto")
print(type(model).__name__)

load_quantized_tribev2() keeps data.video_feature.image.model_name aligned with facebook/vjepa2-vitg-fpc64-256 for compatibility and injects the embedded quantized weights from quantized_video/.

Based on

• facebook/tribev2
• facebook/vjepa2-vitg-fpc64-256