VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning

Overview

Native visual reasoning, i.e., reasoning through visual generation, has recently emerged as a promising direction for studying visual intelligence beyond language. Yet progress remains bottlenecked by the lack of scalable training tasks, reliable feedback, and controlled comparisons across generative substrates. In this work, we introduce VBVR-Pro, a closed-loop testbed that makes native visual reasoning through generation trainable, verifiable, optimizable, and experimentally controllable. 1) Task scaling. VBVR-Pro turns visual reasoning into a controlled task space of 300 procedurally generated tasks. Models trained on VBVR-Pro show strong transfer beyond the proposed suite across six held-out visual reasoning benchmarks such as RISE-Video, MME-CoF-Pro, and BabyVision. Further analysis validates that these gains reflect visual reasoning rather than instruction-pattern fitting. 2) Verifiable rewards. VBVR-Pro provides verifiable reward scorers for task-grounded evaluation. Through a systematic study of leading MLLMs as judges, we identify recurring failure modes of the prevalent VLM-as-a-judge paradigm. In contrast, the proposed scorers are grounded on verifiable task-specific rules, achieve fine-grained alignment with human judgments. Importantly, they serve as reliable reward signals for large-scale multi-task reinforcement learning and demonstrate stronger post-RL performance across visual reasoning tasks. 3) Mechanism study. VBVR-Pro enables controlled modality studies across more than 30 image, video, and interleaved generators. Our analysis shows that video generation remains strongest for tasks requiring persistent spatiotemporal state tracking, while interleaved generation provides a compute-efficient alternative by externalizing intermediate visual states. Critically, ablations and probing confirm the presence of vision-native trajectories, that are a more crucial substrate than explicit linguistic chains of thought for visual reasoning. We release all data, models, scorers, and code to facilitate future research.

The models are presented in the paper VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning.

Models Zoo

Release Information

This repository contains the Hugging Face export of the EMA checkpoint at training step 30,000 from neo_old_script_vbvr_pro. The model takes an initial image and a text instruction and generates one or more sequential keyframes.

The checkpoint uses custom Transformers code. Loading it therefore requires trust_remote_code=True. Review the Python files in this repository before loading code from an untrusted copy.

In this release, we present all models presented in paper VBVR-Pro-Trained-Models, VBVR-Pro-Dataset-Video, VBVR-Pro-Dataset-Image, VBVR-Pro-Bench, VBVR-Pro-Code and VBVR-Bench-Leaderboard.

VBVR-Pro Benchmark Results

Quick Start

Method 1: Standalone Transformers inference

Requirements

• Linux and a CUDA-capable GPU
• Python 3.10 or newer
• PyTorch 2.5 or newer, with a matching TorchVision build
• Transformers 4.51 or newer (Transformers 5 is not currently supported)

Install PyTorch and TorchVision for your CUDA version first, then install the remaining packages:

pip install "transformers>=4.51,<5" safetensors pillow numpy pandas

The checkpoint contains about 36 GB of bfloat16 weights. An 80 GB-class GPU is recommended for straightforward single-GPU inference; activation memory depends strongly on output resolution and the number of generated frames.

Run inference

Download this repository or pass its Hugging Face repository ID directly to the included script:

python inference.py \
  --model Video-Reason/VBVR-Pro-SenseNova-U1 \
  --input first_frame.png \
  --prompt "Move the object to the requested destination while preserving the scene." \
  --num-images 3 \
  --width 512 \
  --height 512 \
  --output-dir outputs

This writes frame_1.png, frame_2.png, and frame_3.png under outputs/. Both output dimensions must be positive multiples of 32.

The equivalent core API is:

import numpy as np
import torch
from PIL import Image
from transformers import AutoModel, AutoTokenizer

model_id = "Video-Reason/VBVR-Pro-SenseNova-U1"
device = "cuda:0"

torch.manual_seed(42)
torch.cuda.manual_seed_all(42)

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModel.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
).to(device).eval()

first_frame = Image.open("first_frame.png").convert("RGB")
num_images = 3

with torch.inference_mode():
    frames = model.interleave_gen_image_only(
        tokenizer,
        "Move the object to the requested destination while preserving the scene.",
        gt_text="<image>" * num_images,
        images=[first_frame],
        image_size=(512, 512),  # (width, height)
        max_images=num_images,
        num_steps=50,
        cfg_scale=1.0,
        img_cfg_scale=1.0,
        timestep_shift=1.0,
    )

for index, frame in enumerate(frames, start=1):
    image = (frame.float() * 0.5 + 0.5).clamp(0, 1)
    array = (
        image[0].permute(1, 2, 0).cpu().numpy() * 255.0
    ).round().astype(np.uint8)
    Image.fromarray(array).save(f"frame_{index}.png")

gt_text controls how many image slots are generated: use one <image> token per requested output frame. Generated frames are sequential: each generated frame is encoded back into the context before the next frame is produced.

Method 2: Unified VBVR-Pro inference

Clone Video-Reason/VBVR-Pro and create its unified inference environment:

git clone https://github.com/Video-Reason/VBVR-Pro.git
cd VBVR-Pro/
uv sync --extra cu124 # or one of [cu118|cu121|cu124|cu126|cu128|cu129]
source .venv/bin/activate

Then run the unified inference script:

python example.py \
  --model_path Video-Reason/VBVR-Pro-SenseNova-U1 \
  --image_paths first_frame.png \
  --prompt "Show the next three steps of the action." \
  --num_images 3 --width 512 --height 512 \
  --output outputs

Settings used by the existing VBVR-Pro evaluator

The previous evaluation path loads this EMA export with AutoModel and AutoTokenizer, then calls interleave_gen_image_only with the following defaults:

For benchmark evaluation, the evaluator removes literal <image> placeholders from the prompt, uses the reference keyframe dimensions after resizing them to multiples of 32, and requests one output image per reference keyframe. The included CLI exposes the same generation API but uses one explicit output size for all frames.

Notes

• This is a custom Neo-Unify image-generation checkpoint, not a Diffusers or Wan checkpoint.
• The model is intended for bfloat16 CUDA inference. CPU inference is not supported by the included script.
• Higher resolutions and additional output frames increase runtime and memory use substantially.
• Use generated content responsibly and in accordance with the model license.

License

This model is released under the Apache License 2.0. Consult the base model repository for any additional terms and notices.

Citation

@misc{xu2026vbvrproscalableverifiablesuite,
      title={VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning}, 
      author={Junxiang Xu and Ruisi Wang and Fanyi Pu and Maijunxian Wang and Ran Ji and Tongxi Zhou and Chenyang Gu and Jing Zuo and Hongcan Xiao and Yimeng Geng and Wanqi Yin and Wei Chen and Oscar Qian and Zhengan Yan and Ziqi Huang and Haiwen Diao and Liang Pan and Bo Li and Xiangyu Fan and Dezhi Luo and Fengyuan Yu and Zehong Zhao and Qingying Gao and Tinghui Zhu and Yilan Zhang and Jingqi Tong and Pinyuan Feng and Zhengze Jiang and Letian Wang and Ziyu Guo and Renrui Zhang and Jieneng Chen and Sonia Joseph and Constantin Venhoff and Saman Motamed and Mengyue Yang and Chandra Sripada and Alan Yuille and Philip Torr and Lvmin Zhang and Vikash Kumar and Daniel Khashabi and Nikolaus Kriegeskorte and Raphaël Millière and Vincent C. Müller and Anyi Rao and Quan Wang and Ziwei Liu and Dahua Lin and Lei Yang and Hokin Deng and Zhongang Cai},
      year={2026},
      eprint={2608.26105},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2608.26105}, 
}