--- library_name: transformers pipeline_tag: image-to-image tags: - custom_code - image-generation - interleaved-generation - vbvr-pro - qwen3 --- # 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](https://huggingface.co/papers/2602.20159). ## Models Zoo

Model	Base Architecture	Other Remarks
Image Generation Models
VBVR-Pro-BAGEL	BAGEL-7B-MoT	Complete model
VBVR-Pro-FLUX2-dev	FLUX.2-dev	Complete model, Diffusers format
VBVR-Pro-FLUX2-dev-diffsynth	FLUX.2-dev	LoRA model, DiffSynth format
VBVR-Pro-Qwen-Image-Edit	Qwen-Image-Edit-2511	Complete model, Diffusers format
VBVR-Pro-Qwen-Image-Edit-diffsynth	Qwen-Image-Edit-2511	LoRA model, DiffSynth format
Interleaved Image Generation Models
VBVR-Pro-ThinkMorph	ThinkMorph-7B	Complete model
VBVR-Pro-SenseNova-U1	SenseNova-U1-8B-MoT	Complete model
Video Generation Models
VBVR-Pro-LTX2.3	LTX-Video-2.3	Complete model, Diffusers format
VBVR-Pro-LTX2.3-diffsynth	LTX-Video-2.3	LoRA model, DiffSynth format
VBVR-Pro-Wan2.1-I2V-14B	Wan2.1-I2V-14B-720P	Complete model, Diffusers format
VBVR-Pro-Wan2.1-I2V-14B-diffsynth	Wan2.1-I2V-14B-720P	LoRA model, DiffSynth format
VBVR-Pro-Wan2.2-I2V-A14B	Wan2.2-I2V-A14B	Complete model, Diffusers format
VBVR-Pro-Wan2.2-I2V-A14B-diffsynth	Wan2.2-I2V-A14B	LoRA model, DiffSynth format
VBVR-Pro-Wan2.2-TI2V-5B	Wan2.2-TI2V-5B	Complete model, Diffusers format
VBVR-Pro-Wan2.2-TI2V-5B-diffsynth	Wan2.2-TI2V-5B	LoRA model, DiffSynth format

## 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**](https://huggingface.co/collections/Video-Reason/VBVR-Pro), [**VBVR-Pro-Dataset-Video**](https://huggingface.co/datasets/Video-Reason/VBVR-Pro-SFT-Video), [**VBVR-Pro-Dataset-Image**](https://huggingface.co/datasets/Video-Reason/VBVR-Pro-SFT-Image), [**VBVR-Pro-Bench**](https://huggingface.co/datasets/Video-Reason/VBVR-Pro-Bench), [**VBVR-Pro-Code**](https://github.com/Video-Reason/VBVR-Pro) and [**VBVR-Bench-Leaderboard**](https://video-reason.com/pro/bench/#leaderboard). ## VBVR-Pro Benchmark Results

Models	Overall	In-Domain by Category						Out-of-Domain by Category
Models	Overall	Avg.	Abst.	Know.	Perc.	Spat.	Trans.	Avg.	Abst.	Know.	Perc.	Spat.	Trans.
Image Generation Models
Proprietary Models
Qwen-Image-2.0	0.313	0.248	0.269	0.196	0.225	0.170	0.132	0.378	0.341	0.235	0.391	0.384	0.080
Seedream-5.0-Pro	0.557	0.485	0.518	0.312	0.509	0.401	0.217	0.629	0.507	0.455	0.661	0.559	0.202
Open-source Models
BAGEL-7B-MoT	0.089	0.066	0.039	0.085	0.067	0.046	0.027	0.111	0.201	0.031	0.073	0.028	0.121
FLUX.2-dev	0.157	0.108	0.088	0.109	0.072	0.100	0.066	0.206	0.197	0.165	0.184	0.241	0.077
Qwen-Image-Edit	0.134	0.108	0.092	0.082	0.100	0.109	0.056	0.159	0.176	0.063	0.141	0.182	0.082
Strong Baselines
VBVR-Pro-BAGEL	0.172	0.168	0.199	0.105	0.110	0.213	0.055	0.176	0.254	0.104	0.148	0.015	0.145
VBVR-Pro-FLUX.2	0.407	0.484	0.483	0.323	0.367	0.449	0.336	0.330	0.361	0.272	0.255	0.454	0.128
VBVR-Pro-Qwen-Image	0.322	0.332	0.298	0.217	0.193	0.431	0.222	0.311	0.341	0.239	0.233	0.413	0.181
Interleaved Image Generation Models
Proprietary Models
GPT-Image-2	0.507	0.428	0.456	0.318	0.428	0.206	0.300	0.587	0.398	0.413	0.633	0.480	0.303
Nano Banana Pro	0.564	0.480	0.518	0.422	0.512	0.285	0.174	0.648	0.553	0.499	0.657	0.585	0.220
Open-source Models
ThinkMorph-7B	0.154	0.113	0.100	0.082	0.101	0.148	0.031	0.195	0.176	0.166	0.163	0.253	0.103
VBVR-SenseNova-U1	0.408	0.469	0.356	0.313	0.373	0.386	0.477	0.347	0.291	0.317	0.275	0.480	0.238
SenseNova-U1-8B-MoT	0.565	0.533	0.501	0.395	0.544	0.355	0.349	0.597	0.448	0.495	0.533	0.717	0.401
Strong Baselines
VBVR-Pro-ThinkMorph	0.373	0.402	0.403	0.344	0.238	0.454	0.184	0.344	0.367	0.224	0.238	0.535	0.257
VBVR-Pro-SenseNova-U1	0.638	0.811	0.648	0.695	0.621	0.770	0.541	0.464	0.480	0.328	0.344	0.558	0.408
Video Generation Models
Proprietary Models
Veo 3.1	0.309	0.312	0.275	0.299	0.252	0.267	0.157	0.305	0.305	0.233	0.252	0.312	0.219
Kling V3	0.392	0.356	0.213	0.326	0.320	0.355	0.229	0.427	0.294	0.564	0.375	0.242	0.412
SeedDance 2.0	0.499	0.451	0.338	0.361	0.353	0.468	0.308	0.547	0.369	0.511	0.478	0.538	0.532
Open-source Models
HunyuanVideo-I2V	0.054	0.054	0.023	0.064	0.015	0.084	0.032	0.053	0.088	0.014	0.028	0.062	0.055
CogVideoX1.5-5B-I2V	0.085	0.100	0.061	0.118	0.069	0.092	0.060	0.070	0.125	0.038	0.051	0.040	0.024
Wan2.1-I2V-14B	0.100	0.105	0.052	0.125	0.091	0.102	0.052	0.095	0.112	0.073	0.071	0.123	0.044
Wan2.2-TI2V-5B	0.094	0.066	0.029	0.073	0.050	0.083	0.031	0.122	0.156	0.052	0.106	0.063	0.099
Wan2.2-I2V-14B-720P	0.182	0.157	0.082	0.131	0.110	0.161	0.156	0.207	0.224	0.139	0.140	0.195	0.273
LTX2.3-I2AV	0.112	0.106	0.062	0.109	0.070	0.133	0.055	0.119	0.161	0.135	0.086	0.091	0.050
VBVR-Wan2.2	0.517	0.548	0.237	0.499	0.334	0.566	0.591	0.486	0.310	0.343	0.345	0.732	0.684
Strong Baselines
VBVR-Pro-LTX2.3	0.425	0.527	0.409	0.510	0.346	0.460	0.390	0.324	0.381	0.108	0.201	0.477	0.386
VBVR-Pro-Wan2.1-I2V-14B	0.562	0.730	0.617	0.580	0.452	0.676	0.623	0.395	0.410	0.305	0.230	0.617	0.439
VBVR-Pro-Wan2.2-TI2V-5B	0.470	0.641	0.528	0.556	0.373	0.565	0.557	0.300	0.333	0.127	0.161	0.505	0.409
VBVR-Pro-Wan2.2-I2V-14B	0.670	0.808	0.632	0.685	0.556	0.751	0.636	0.532	0.479	0.418	0.350	0.679	0.690

## 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: ```bash 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: ```bash 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: ```python 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="" * 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 `` 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](https://github.com/Video-Reason/VBVR-Pro) and create its unified inference environment: ```bash 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: ```bash 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: | Setting | Value | | --- | --- | | Denoising steps | 50 | | Text CFG scale | 1.0 | | Image CFG scale | 1.0 | | Timestep shift | 1.0 | | Seed | 42 | | Input | `first_frame.png` plus `prompt.txt` | | Output count | Number of reference `frame_N.png` files | For benchmark evaluation, the evaluator removes literal `` 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 follow the terms that accompany the eventual Hugging Face repository release.