---
language:
- en
base_model:
- Yuppie1204/NeoVerse
tags:
- cornerbench
- video-generation
- multi-view
- autonomous-driving
- control-branch
---

# CornerFactory: NeoObj, NeoView, and NeoEmb

This repository groups the three task-specific control branches used by CornerFactory. The release uses the paper names **NeoObj**, **NeoView**, and **NeoEmb**, with one subdirectory per module.

This release provides the selected **1280 × 720** control-branch weights. They are not the complete NeoVerse backbone, and they are not interchangeable with a standalone Diffusers pipeline checkpoint.

| Module | Task | File |
| --- | --- | --- |
| NeoObj | Masked cross-view object completion | `NeoObj/model.safetensors` |
| NeoView | Perturbed-pose novel-view completion | `NeoView/model.safetensors` |
| NeoEmb | Vehicle-configuration transfer | `NeoEmb/model.safetensors` |

Each file contains 240 BF16 tensors, with 3,049,216,192 parameters in the control branch (approximately 6.10 GB per module). All three training configurations use 720 × 1280 images and 81-frame clips. This does not mean that every released example has 81 frames.

## Download

```python
from huggingface_hub import hf_hub_download

checkpoint = hf_hub_download(
    repo_id="cornerbench-anon-2027/CornerFactory",
    filename="NeoObj/model.safetensors",  # or NeoView / NeoEmb
)
```

During private review, the account that owns the repository must authenticate before downloading. Once public, downloading these files will not require an application or access request.

## Loading the control branch

The matching CornerFactory/NeoVerse pipeline must already be initialized with its backbone, text encoder, tokenizer, VAE, and a compatible controller architecture. The files in this repository replace the controller parameters only.

```python
from load_control_branch import load_control_branch

# `pipe` is an already-initialized compatible pipeline.
load_control_branch(pipe.control_branch, checkpoint)
```

`load_control_branch.py` checks all parameter names and shapes and loads strictly. It does not reconstruct the complete inference pipeline. Task-specific camera calibration, conditioning, preprocessing, scheduler settings, and any inference acceleration LoRA must match the accompanying implementation. End-to-end inference code is a separate release artifact.

## Evaluation scope

These 720p weights are provided for inspection and subsequent reproduction. Numerical results depend on the corresponding evaluation configuration; this release does not claim that these weights reproduce every result in the paper.

## Integrity

See `manifest.json`, each module's `release_config.json`, and `SHA256SUMS`. All checkpoint tensors were checked for finite values, and the release files match the source checkpoint hashes. This integrity check is not an end-to-end inference benchmark.

## Related data

The separate [CornerBench-1200 dataset](https://huggingface.co/datasets/cornerbench-anon-2027/CornerBench-1200) contains three-view generated driving cases. These generated observations are research artifacts and do not establish real-world driving safety.