Professor in Central South University， NUS PHD. Focus on Multi-modality Learning and Data-centric AI.
Singapore
Joined April 2017
- Tweets191
- Following176
- Followers576
- Likes389
Pinned Tweet
Humans see text — but LLMs don’t.
I wrote a short blog post exploring how models can perceive text visually rather than tokenize it:
🔗 csu-jpg.github.io/Blog/peopl…
From PIXEL, CLIPPO, VisInContext, VIST to DeepSeek-OCR, this is a quick story of how vision-centric modeling is changing how machines read,
and a reflection on some of our own small efforts in the past two years.
Excited to share 3 papers at #ECCV2026 🎉
🔹 FlowInOne: unifying multimodal generation as image-in, image-out flow matching
📅 Fri 9/11, 16:00–18:00 CEST | ExHall #149
🔹 Glance: one sample distillation model
📅 Sat 9/12, 10:30–12:30 CEST | ExHall #38
🔹 Evaluating and enhancing negation comprehension in remote sensing MLLMs
Come by the posters and chat, would love to talk vision-centric MLLMs, multimodal generation, and remote sensing understanding!
🧵 ICML 2026 (Seoul) — come say hi! 4 papers from our lab this week:
📌 Today (Jul 8)
14:30 KST, Hall A #314 — TextAtlas5M: a large-scale dataset for dense text image generation
10:45-11:00 KST Oral / poster 14:30-16:15, Hall A #2800 — When the Prompt Becomes Visual: Vision-Centric Jailbreak Attacks for Large Image Editing Models
📌 Tomorrow (Jul 9)
10:30-12:15 KST, Hall A #314 — Entropy-Aware Dynamic KV Cache Sparsification for Autoregressive Image Generation and Editing
9:30-11:15 KST, Hall A #1302 — Mitigating Error Accumulation in Continuous Navigation via Memory-Augmented Kalman Filtering
Spanning text-rich generation, jailbreak/defense, efficient AR generation, and embodied navigation — would love to chat about any of it. Swing by! 🙌
#ICML2026 #MLLM #ImageGeneration
Jinpeng Wang retweeted
can AI write engaging news that people can trust?
introducing ✨Data2Story: a data journalist agent.
give it raw data, it generate a verifiable, multimodal article.
🔍verifiable: every claim is evidence-grounded, traces back to data, code, or a cited source.
🔮multimodal: the article is a generative UI — images, videos, audio, interactive charts.
not just readable, but trustworthy and playable. 🧵1/N
VJA： ICML 2026 Oral 🚨
Modern image editors understand your drawings, arrows, and scribbles — not just text. We asked: can an attacker use that too?
Yes. Introducing VJA: hide a malicious instruction inside the image itself, bypass text-based safety filters entirely.
Up to 80.9% attack success on SOTA commercial models. We also built a benchmark (IESBench) + a training-free defense.
Paper: arxiv.org/abs/2602.10179
#icml #imageediting #generation
Jinpeng Wang retweeted
🔥 Multimodal gen must be text-driven?
Meet FlowInOne 🖼️➡️🖼️: Image-in, Image-out
1️⃣ Vision-Centric: no text encoders or bridges.
2️⃣ Continuous evolution: visual Flow Matching.
3️⃣ VisPrompt-5M: across diverse I2I tasks.
📄 arxiv.org/abs/2604.06757
🌐 csu-jpg.github.io/FlowInOne.…
Adapt the Qwen-Image model to a target domain with one single example.
Thanks AK for sharing our work. We speed up Qwen-image and Flux
Inference speed from 50 steps
to less than 10steps. 1 sample is enough for speeding up diffusion model in specific domain.
Paper Link: arxiv.org/abs/2512.02899
Code Link: github.com/CSU-JPG/Glance
Thanks AK for sharing our work. We speed up Qwen-image and Flux
Inference speed from 50 steps
to less than 10steps. 1 sample is enough for speeding up diffusion model in specific domain.
Paper Link: arxiv.org/abs/2512.02899
Code Link: github.com/CSU-JPG/Glance
Jinpeng Wang retweeted
We added Gemini3-Pro #Gemini results on VCode!
Check out at csu-jpg.github.io/VCode/
🏅Gemini3-Pro performs quite well at image2svg.
It achieved 52.1 in VCode, surpassing GPT-5's 46.8 (+5.3).
Open-sourced at github.com/CSU-JPG/VCode
thanks JIQIZHIXIN for sharing our work. We use visual encoder to compress long text.
As LLMs grow to trillions of parameters and context windows stretch to hundreds of thousands of tokens, compute costs explode.
Vision Centric Token Compression (Vist) offers a fix — a vision-centric token compression method that mimics human reading.
It converts distant context into images for a vision encoder to skim, while the LLM focuses on nearby text.
Matching full-context accuracy with 2.3× fewer tokens, Vist slashes FLOPs by 16%, memory by 50%, and outperforms prior compression methods by 7.6% across major benchmarks.
Vision-centric Token Compression in Large Language Model
NUST, CSU, NFU
Paper: arxiv.org/abs/2502.00791
Our report: mp.weixin.qq.com/s/zYnxpBhRs…
📬 #PapersAccepted by Jiqizhixin
🤣🤣🤣🤣🤣
😲Turn any image into symbolic SVG at
huggingface.co/spaces/CSU-JP…
💻Github: github.com/CSU-JPG/VCode
Built with @xinlinzyh and @zzranhangyu
Thanks @_akhaliq for sharing and Gradio demo team
@huggingface @Gradio!
Jinpeng Wang retweeted
Thanks @_akhaliq sharing our work!
🚀 Glad to introduce our newest work — VCode!
🎨 VCode: A Multimodal Coding Benchmark with SVG as Symbolic Visual Representation
For decades, RGB pixels have been the default medium for representing images. But in the agentic era, how can we move beyond raw pixels toward interpretable, executable, and symbolic visual representations?
🔍 We address this question with VCode, which reframes visual representation as SVG code — aligning with how humans reason over sketches and symbolic abstractions.
👥 Happy to collaborate with: Yuhao Zheng, Hangyu Ran, Dantong Zhu, Dongxing Mao, @LINJIEFUN , @philiptorr and @awinyimgprocess
📄 arXiv: arxiv.org/pdf/2511.02778
🌐 website: csu-jpg.github.io/VCode/
💻 github: github.com/CSU-JPG/VCode
🤗 Huggingface daily paper: huggingface.co/papers/2511.0…
"Huamn mnid deos not raed ervey lteter by istlef, but the wrod as a wlohes (Rawlinson, 1976)." – Graham Rawlinson
🚀 We propose new work: “See the Text: From Tokenization to Visual Reading” ，which renders text as images and uses a vision-centric pipeline instead of traditional subword tokenization.
• Up to 4.4× fewer tokens & ~70% less FLOPs
• Better cross-lingual transfer & robustness to typos/surface noise
• Works on not only long-text compression, but also VQA and classification in NLP.
👉 arXiv: arxiv.org/pdf/2510.18840
