Modelpublic

Video-Reason/VBVR-Pro-ThinkMorph

sourceHugging Faceapache-2.0updated 26d agoView on Hugging Face

0likes42downloads

Model Card

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

<table border="1" cellspacing="0" cellpadding="4" style="border-collapse: collapse; width: 100%;"> <thead> <tr> <th width="260" style="min-width: 260px;">Model</th> <th>Base Architecture</th> <th>Other Remarks</th> </tr> </thead> <tbody> <tr style="border-top: 4px solid #6b7280; background-color: #e5e7eb;"><th colspan="3" align="left">Image Generation Models</th></tr> <tr><td><a href="https://huggingface.co/Video-Reason/VBVR-Pro-BAGEL">VBVR-Pro-BAGEL</a></td><td>BAGEL-7B-MoT</td><td>Complete model</td></tr> <tr><td><a href="https://huggingface.co/Video-Reason/VBVR-Pro-FLUX2-dev">VBVR-Pro-FLUX2-dev</a></td><td>FLUX.2-dev</td><td>Complete model, Diffusers format</td></tr> <tr><td><a href="https://huggingface.co/Video-Reason/VBVR-Pro-FLUX2-dev-diffsynth">VBVR-Pro-FLUX2-dev-diffsynth</a></td><td>FLUX.2-dev</td><td>LoRA model, DiffSynth format</td></tr> <tr><td><a href="https://huggingface.co/Video-Reason/VBVR-Pro-Qwen-Image-Edit">VBVR-Pro-Qwen-Image-Edit</a></td><td>Qwen-Image-Edit-2511</td><td>Complete model, Diffusers format</td></tr> <tr><td><a href="https://huggingface.co/Video-Reason/VBVR-Pro-Qwen-Image-Edit-diffsynth">VBVR-Pro-Qwen-Image-Edit-diffsynth</a></td><td>Qwen-Image-Edit-2511</td><td>LoRA model, DiffSynth format</td></tr> <tr style="border-top: 4px solid #6b7280; background-color: #e5e7eb;"><th colspan="3" align="left">Interleaved Image Generation Models</th></tr> <tr><td><a href="https://huggingface.co/Video-Reason/VBVR-Pro-ThinkMorph">VBVR-Pro-ThinkMorph</a></td><td>ThinkMorph-7B</td><td>Complete model</td></tr> <tr><td><a href="https://huggingface.co/Video-Reason/VBVR-Pro-SenseNova-U1">VBVR-Pro-SenseNova-U1</a></td><td>SenseNova-U1-8B-MoT</td><td>Complete model</td></tr> <tr style="border-top: 4px solid #6b7280; background-color: #e5e7eb;"><th colspan="3" align="left">Video Generation Models</th></tr> <tr><td><a href="https://huggingface.co/Video-Reason/VBVR-Pro-LTX2.3">VBVR-Pro-LTX2.3</a></td><td>LTX-Video-2.3</td><td>Complete model, Diffusers format</td></tr> <tr><td><a href="https://huggingface.co/Video-Reason/VBVR-Pro-LTX2.3-diffsynth">VBVR-Pro-LTX2.3-diffsynth</a></td><td>LTX-Video-2.3</td><td>LoRA model, DiffSynth format</td></tr> <tr><td><a href="https://huggingface.co/Video-Reason/VBVR-Pro-Wan2.1-I2V-14B">VBVR-Pro-Wan2.1-I2V-14B</a></td><td>Wan2.1-I2V-14B-720P</td><td>Complete model, Diffusers format</td></tr> <tr><td><a href="https://huggingface.co/Video-Reason/VBVR-Pro-Wan2.1-I2V-14B-diffsynth">VBVR-Pro-Wan2.1-I2V-14B-diffsynth</a></td><td>Wan2.1-I2V-14B-720P</td><td>LoRA model, DiffSynth format</td></tr> <tr><td><a href="https://huggingface.co/Video-Reason/VBVR-Pro-Wan2.2-I2V-A14B">VBVR-Pro-Wan2.2-I2V-A14B</a></td><td>Wan2.2-I2V-A14B</td><td>Complete model, Diffusers format</td></tr> <tr><td><a href="https://huggingface.co/Video-Reason/VBVR-Pro-Wan2.2-I2V-A14B-diffsynth">VBVR-Pro-Wan2.2-I2V-A14B-diffsynth</a></td><td>Wan2.2-I2V-A14B</td><td>LoRA model, DiffSynth format</td></tr> <tr><td><a href="https://huggingface.co/Video-Reason/VBVR-Pro-Wan2.2-TI2V-5B">VBVR-Pro-Wan2.2-TI2V-5B</a></td><td>Wan2.2-TI2V-5B</td><td>Complete model, Diffusers format</td></tr> <tr><td><a href="https://huggingface.co/Video-Reason/VBVR-Pro-Wan2.2-TI2V-5B-diffsynth">VBVR-Pro-Wan2.2-TI2V-5B-diffsynth</a></td><td>Wan2.2-TI2V-5B</td><td>LoRA model, DiffSynth format</td></tr> </tbody> </table>

Release Information

This repository contains a VBVR-Pro fine-tune of ThinkMorph-7B. It takes an initial image and a text instruction, reasons with interleaved text and images, and can generate one or more sequential images before returning its final answer.

The released weights are the EMA checkpoint from training step 25,000. The training checkpoint was stored in FP32; this Hugging Face export is converted to BF16, matching the precision of the original ThinkMorph-7B release.

Repository contents

File	Purpose
`ema.safetensors`	Complete BF16 fine-tuned ThinkMorph model state
`ae.safetensors`	Image autoencoder used for encoding and generation
`llm_config.json`	Qwen2-MoT language-model configuration
`vit_config.json`	SigLIP vision-encoder configuration
`tokenizer.json`, `tokenizer_config.json`, `vocab.json`, `merges.txt`	Tokenizer assets
`generation_config.json`	Default text-generation settings
`config.json`	Checkpoint family identifier

This is a full checkpoint, not a LoRA or other adapter. No separate base-model download is needed after this repository has been downloaded.

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

<table border="1" cellspacing="0" cellpadding="4" style="border-collapse: collapse; width: 100%; font-size: 12px;"> <thead> <tr> <th rowspan="2" width="260" style="min-width: 260px;">Models</th> <th rowspan="2">Overall</th> <th colspan="6">In-Domain by Category</th> <th colspan="6">Out-of-Domain by Category</th> </tr> <tr> <th>Avg.</th><th>Abst.</th><th>Know.</th><th>Perc.</th><th>Spat.</th><th>Trans.</th> <th>Avg.</th><th>Abst.</th><th>Know.</th><th>Perc.</th><th>Spat.</th><th>Trans.</th> </tr> </thead> <tbody> <tr style="border-top: 4px solid #6b7280; background-color: #e5e7eb;"><th colspan="14" align="left">Image Generation Models</th></tr> <tr style="background-color: #dbeafe;"><th colspan="14" align="left">Proprietary Models</th></tr> <tr><td>Qwen-Image-2.0</td><td>0.313</td><td>0.248</td><td>0.269</td><td>0.196</td><td>0.225</td><td>0.170</td><td>0.132</td><td>0.378</td><td>0.341</td><td>0.235</td><td>0.391</td><td>0.384</td><td>0.080</td></tr> <tr><td>Seedream-5.0-Pro</td><td>0.557</td><td>0.485</td><td>0.518</td><td>0.312</td><td>0.509</td><td>0.401</td><td>0.217</td><td>0.629</td><td>0.507</td><td>0.455</td><td>0.661</td><td>0.559</td><td>0.202</td></tr> <tr style="background-color: #dbeafe;"><th colspan="14" align="left">Open-source Models</th></tr> <tr><td>BAGEL-7B-MoT</td><td>0.089</td><td>0.066</td><td>0.039</td><td>0.085</td><td>0.067</td><td>0.046</td><td>0.027</td><td>0.111</td><td>0.201</td><td>0.031</td><td>0.073</td><td>0.028</td><td>0.121</td></tr> <tr><td>FLUX.2-dev</td><td>0.157</td><td>0.108</td><td>0.088</td><td>0.109</td><td>0.072</td><td>0.100</td><td>0.066</td><td>0.206</td><td>0.197</td><td>0.165</td><td>0.184</td><td>0.241</td><td>0.077</td></tr> <tr><td>Qwen-Image-Edit</td><td>0.134</td><td>0.108</td><td>0.092</td><td>0.082</td><td>0.100</td><td>0.109</td><td>0.056</td><td>0.159</td><td>0.176</td><td>0.063</td><td>0.141</td><td>0.182</td><td>0.082</td></tr> <tr style="background-color: #dbeafe;"><th colspan="14" align="left">Strong Baselines</th></tr> <tr><td>VBVR-Pro-BAGEL</td><td>0.172</td><td>0.168</td><td>0.199</td><td>0.105</td><td>0.110</td><td>0.213</td><td>0.055</td><td>0.176</td><td>0.254</td><td>0.104</td><td>0.148</td><td>0.015</td><td>0.145</td></tr> <tr><td>VBVR-Pro-FLUX.2</td><td>0.407</td><td>0.484</td><td>0.483</td><td>0.323</td><td>0.367</td><td>0.449</td><td>0.336</td><td>0.330</td><td>0.361</td><td>0.272</td><td>0.255</td><td>0.454</td><td>0.128</td></tr> <tr><td>VBVR-Pro-Qwen-Image</td><td>0.322</td><td>0.332</td><td>0.298</td><td>0.217</td><td>0.193</td><td>0.431</td><td>0.222</td><td>0.311</td><td>0.341</td><td>0.239</td><td>0.233</td><td>0.413</td><td>0.181</td></tr> <tr style="border-top: 4px solid #6b7280; background-color: #e5e7eb;"><th colspan="14" align="left">Interleaved Image Generation Models</th></tr> <tr style="background-color: #dbeafe;"><th colspan="14" align="left">Proprietary Models</th></tr> <tr><td>GPT-Image-2</td><td>0.507</td><td>0.428</td><td>0.456</td><td>0.318</td><td>0.428</td><td>0.206</td><td>0.300</td><td>0.587</td><td>0.398</td><td>0.413</td><td>0.633</td><td>0.480</td><td>0.303</td></tr> <tr><td>Nano Banana Pro</td><td>0.564</td><td>0.480</td><td>0.518</td><td>0.422</td><td>0.512</td><td>0.285</td><td>0.174</td><td>0.648</td><td>0.553</td><td>0.499</td><td>0.657</td><td>0.585</td><td>0.220</td></tr> <tr style="background-color: #dbeafe;"><th colspan="14" align="left">Open-source Models</th></tr> <tr><td>ThinkMorph-7B</td><td>0.154</td><td>0.113</td><td>0.100</td><td>0.082</td><td>0.101</td><td>0.148</td><td>0.031</td><td>0.195</td><td>0.176</td><td>0.166</td><td>0.163</td><td>0.253</td><td>0.103</td></tr> <tr><td>VBVR-SenseNova-U1</td><td>0.408</td><td>0.469</td><td>0.356</td><td>0.313</td><td>0.373</td><td>0.386</td><td>0.477</td><td>0.347</td><td>0.291</td><td>0.317</td><td>0.275</td><td>0.480</td><td>0.238</td></tr> <tr><td>SenseNova-U1-8B-MoT</td><td>0.565</td><td>0.533</td><td>0.501</td><td>0.395</td><td>0.544</td><td>0.355</td><td>0.349</td><td>0.597</td><td>0.448</td><td>0.495</td><td>0.533</td><td>0.717</td><td>0.401</td></tr> <tr style="background-color: #dbeafe;"><th colspan="14" align="left">Strong Baselines</th></tr> <tr><td>VBVR-Pro-ThinkMorph</td><td>0.373</td><td>0.402</td><td>0.403</td><td>0.344</td><td>0.238</td><td>0.454</td><td>0.184</td><td>0.344</td><td>0.367</td><td>0.224</td><td>0.238</td><td>0.535</td><td>0.257</td></tr> <tr><td>VBVR-Pro-SenseNova-U1</td><td>0.638</td><td>0.811</td><td>0.648</td><td>0.695</td><td>0.621</td><td>0.770</td><td>0.541</td><td>0.464</td><td>0.480</td><td>0.328</td><td>0.344</td><td>0.558</td><td>0.408</td></tr> <tr style="border-top: 4px solid #6b7280; background-color: #e5e7eb;"><th colspan="14" align="left">Video Generation Models</th></tr> <tr style="background-color: #dbeafe;"><th colspan="14" align="left">Proprietary Models</th></tr> <tr><td>Veo 3.1</td><td>0.309</td><td>0.312</td><td>0.275</td><td>0.299</td><td>0.252</td><td>0.267</td><td>0.157</td><td>0.305</td><td>0.305</td><td>0.233</td><td>0.252</td><td>0.312</td><td>0.219</td></tr> <tr><td>Kling V3</td><td>0.392</td><td>0.356</td><td>0.213</td><td>0.326</td><td>0.320</td><td>0.355</td><td>0.229</td><td>0.427</td><td>0.294</td><td>0.564</td><td>0.375</td><td>0.242</td><td>0.412</td></tr> <tr><td>SeedDance 2.0</td><td>0.499</td><td>0.451</td><td>0.338</td><td>0.361</td><td>0.353</td><td>0.468</td><td>0.308</td><td>0.547</td><td>0.369</td><td>0.511</td><td>0.478</td><td>0.538</td><td>0.532</td></tr> <tr style="background-color: #dbeafe;"><th colspan="14" align="left">Open-source Models</th></tr> <tr><td>HunyuanVideo-I2V</td><td>0.054</td><td>0.054</td><td>0.023</td><td>0.064</td><td>0.015</td><td>0.084</td><td>0.032</td><td>0.053</td><td>0.088</td><td>0.014</td><td>0.028</td><td>0.062</td><td>0.055</td></tr> <tr><td>CogVideoX1.5-5B-I2V</td><td>0.085</td><td>0.100</td><td>0.061</td><td>0.118</td><td>0.069</td><td>0.092</td><td>0.060</td><td>0.070</td><td>0.125</td><td>0.038</td><td>0.051</td><td>0.040</td><td>0.024</td></tr> <tr><td>Wan2.1-I2V-14B</td><td>0.100</td><td>0.105</td><td>0.052</td><td>0.125</td><td>0.091</td><td>0.102</td><td>0.052</td><td>0.095</td><td>0.112</td><td>0.073</td><td>0.071</td><td>0.123</td><td>0.044</td></tr> <tr><td>Wan2.2-TI2V-5B</td><td>0.094</td><td>0.066</td><td>0.029</td><td>0.073</td><td>0.050</td><td>0.083</td><td>0.031</td><td>0.122</td><td>0.156</td><td>0.052</td><td>0.106</td><td>0.063</td><td>0.099</td></tr> <tr><td>Wan2.2-I2V-14B-720P</td><td>0.182</td><td>0.157</td><td>0.082</td><td>0.131</td><td>0.110</td><td>0.161</td><td>0.156</td><td>0.207</td><td>0.224</td><td>0.139</td><td>0.140</td><td>0.195</td><td>0.273</td></tr> <tr><td>LTX2.3-I2AV</td><td>0.112</td><td>0.106</td><td>0.062</td><td>0.109</td><td>0.070</td><td>0.133</td><td>0.055</td><td>0.119</td><td>0.161</td><td>0.135</td><td>0.086</td><td>0.091</td><td>0.050</td></tr> <tr><td>VBVR-Wan2.2</td><td>0.517</td><td>0.548</td><td>0.237</td><td>0.499</td><td>0.334</td><td>0.566</td><td>0.591</td><td>0.486</td><td>0.310</td><td>0.343</td><td>0.345</td><td>0.732</td><td>0.684</td></tr> <tr style="background-color: #dbeafe;"><th colspan="14" align="left">Strong Baselines</th></tr> <tr><td>VBVR-Pro-LTX2.3</td><td>0.425</td><td>0.527</td><td>0.409</td><td>0.510</td><td>0.346</td><td>0.460</td><td>0.390</td><td>0.324</td><td>0.381</td><td>0.108</td><td>0.201</td><td>0.477</td><td>0.386</td></tr> <tr><td>VBVR-Pro-Wan2.1-I2V-14B</td><td>0.562</td><td>0.730</td><td>0.617</td><td>0.580</td><td>0.452</td><td>0.676</td><td>0.623</td><td>0.395</td><td>0.410</td><td>0.305</td><td>0.230</td><td>0.617</td><td>0.439</td></tr> <tr><td>VBVR-Pro-Wan2.2-TI2V-5B</td><td>0.470</td><td>0.641</td><td>0.528</td><td>0.556</td><td>0.373</td><td>0.565</td><td>0.557</td><td>0.300</td><td>0.333</td><td>0.127</td><td>0.161</td><td>0.505</td><td>0.409</td></tr> <tr><td>VBVR-Pro-Wan2.2-I2V-14B</td><td>0.670</td><td>0.808</td><td>0.632</td><td>0.685</td><td>0.556</td><td>0.751</td><td>0.636</td><td>0.532</td><td>0.479</td><td>0.418</td><td>0.350</td><td>0.679</td><td>0.690</td></tr> </tbody> </table>

Quick Start

Method 1: Standalone ThinkMorph inference

Requirements

—Linux and a CUDA-capable GPU
—Python 3.10
—The ThinkMorph inference code and its dependencies

For straightforward inference, an 80 GB-class GPU is recommended. The ThinkMorph loader can also distribute the model across multiple visible GPUs.

bash

git clone https://github.com/ThinkMorph/ThinkMorph.git
cd ThinkMorph
conda create -n thinkmorph python=3.10 -y
conda activate thinkmorph
pip install -r requirements.txt

Download

python

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="Video-Reason/VBVR-Pro-ThinkMorph",
    local_dir="models/VBVR-Pro-ThinkMorph",
    allow_patterns=["*.json", "*.safetensors", "*.txt", "*.md"],
)

Inference

Run the following from the ThinkMorph repository. The same downloaded directory is passed as both model_path and config_path because this release is self-contained.

python

import random

import numpy as np
import torch
from PIL import Image

from inferencer import InterleaveInferencer
from scripts.cza.inference import load_model

model_path = "models/VBVR-Pro-ThinkMorph"

random.seed(42)
np.random.seed(42)
torch.manual_seed(42)
torch.cuda.manual_seed_all(42)

model, vae, tokenizer, token_ids, vae_transform, vit_transform = load_model(
    model_path,
    model_path,
)
inferencer = InterleaveInferencer(
    model, vae, tokenizer, vae_transform, vit_transform, token_ids
)

system_prompt = (
    "Let's think step by step to answer the question. For text-based thinking, "
    "enclose the process within <think> </think>. For visual thinking, enclose "
    "the content within <image_start> </image_end>. Finally conclude with the "
    "final answer wrapped in <answer> </answer>"
)

input_image = Image.open("first_frame.png").convert("RGB")
output = inferencer(
    image=input_image,
    text="Move the object to the requested destination while preserving the scene.",
    understanding_output=False,
    think=True,
    system_prompt=system_prompt,
    max_think_token_n=4096,
    do_sample=True,
    text_temperature=0.3,
    cfg_text_scale=4.0,
    cfg_img_scale=2.0,
    cfg_interval=[0.0, 1.0],
    timestep_shift=3.0,
    num_timesteps=50,
    cfg_renorm_min=0.0,
    cfg_renorm_type="text_channel",
    max_rounds=10,
)

for index, item in enumerate(output):
    if isinstance(item, Image.Image):
        item.save(f"generated_{index}.png")
    else:
        print(item, end="")

The settings above reproduce the first-attempt settings used by the existing VBVR-Pro evaluator:

Setting	Value
Seed	42
Maximum thinking tokens	4096
Text temperature	0.3
Text CFG scale	4.0
Image CFG scale	2.0
Denoising steps	50
Timestep shift	3.0
Maximum interleaved rounds	10

Method 2: Unified VBVR-Pro inference

Clone 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-ThinkMorph \
  --image_paths first_frame.png \
  --prompt "Solve the task step by step." \
  --think --max_rounds 10 --num_images 2 \
  --output outputs

Notes

—The model uses the custom ThinkMorph/BAGEL architecture and is not directly loadable with transformers.AutoModel or a Diffusers pipeline.
—Inputs should follow the same image-plus-instruction format used by ThinkMorph. Generated images appear as PIL.Image.Image objects interleaved with text strings.
—More generated images, larger resolutions, and longer reasoning traces increase runtime and memory use.
—This checkpoint may inherit limitations and biases from its base model and fine-tuning data. Validate outputs for your application and use generated content responsibly.

Citation

bibtex

@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}, 
}

License

This model is released under the Apache License 2.0. See the base model repository for additional context.