# WithAnyone — Break Free from the AI “Copy-Paste” Curse
**Date:** 2025-11-16 11:59
**Location:** Hong Kong, China
---
## Capture Your Unique AI Photo
Using **WithAnyone**, you can capture your **own unique, personalized photo** — whether alone or with friends.
**WithAnyone** lets you create photos *with anyone, anywhere!*
Jointly developed by **Fudan University** and **Jiepue Xingchen**, this groundbreaking AI group photo generation model breaks the long-standing “copy-paste” limitation. Simply:
1. **Upload** a photo.
2. **Generate** a natural, realistic, seamless AI image — *with one click*.
---
## What Is WithAnyone?
WithAnyone is a **personalized AI photo generation method** that:
- From a **single photo of a person**, creates images with varied angles and facial expressions.
- From **one photo each of multiple people**, places them together into a natural-looking, harmonious group photo.
**Previous methods** — such as **InstantID** and **PuLID** — had a key limitation:
They would **replicate the exact pose and expression** from the input photo, making variation impossible.
**WithAnyone breaks this limitation!**
It achieves **ID consistency** *and* **ID controllability** at leading levels.
---
## How Effective Is It?
Rather than just talk, let’s see the **results**:
**Multi-person example:**
**Large group examples:**
> Three sisters in matching sunglasses; three brothers in suits — naturally posed and blended into the scene.
**Difference from past models:**
Older methods “copy-pasted” expressions directly from reference images.
WithAnyone **preserves identity** while generating realistic group compositions.
> ✅ The **ComfyUI version** is **already online** — workflow included for quick deployment!
---
## How Does WithAnyone Break the Curse?
The team first identified the **root problem**: overly similar AI portraits that *out-match* real-life image variation for the same person.
To measure this, they introduced:
- **MultiID-Bench** benchmark
- **“Copy-paste” metric**
**Metric definition:**
> *Distance from output to reference* **minus** *distance from output to target*.
Findings:
Existing methods produce outputs **too close** to the reference — forming a strong fitted line correlation.
---
### Data & Training Innovation
Past personalized generation lacked **multi-angle data** for the same person and used **reconstruction training** (input reference = supervision target).
**WithAnyone’s approach:**
- Curated **500,000+ multi-person group photos**
Each person matched with hundreds of varied references: angles, expressions, lighting.
- Added **millions** of unmatched group photos.
This is a **weighty open-source release**: **Model + Benchmark + Dataset**.
---
### Training Pipeline
1. **Base training**: traditional reconstruction training.
2. **Paired data training**: link people across varied angles/expressions.
3. **Fine-tuning**: high-quality curated samples.
### Key Innovations
- **Reference image dual-channel encoding**
Face recognition network provides **identity vectors**, image encoder captures **visual features**.
Facial embeddings **apply only to face regions**, preventing cross-face leakage.
- **Contrastive learning**
Same-identity images with varied styles/poses are drawn **closer**, different identities **pushed apart**.
- **Balanced optimization objectives**
Standard diffusion/reconstruction loss + **ID-contrastive loss** + **ground-truth-aligned ID loss**.
This balances **similarity** and **variation**.
---
## Try WithAnyone
WithAnyone is **fully open-sourced** with:
- Code & model weights
- Sample datasets
- Benchmark (**MultiID-Bench**)
- Public links via Hugging Face & GitHub
- ComfyUI support
**Resources:**
- GitHub: [https://github.com/Doby-Xu/WithAnyone](https://github.com/Doby-Xu/WithAnyone)
- ComfyUI repo: `okdalto/ComfyUI-WithAnyone`
- Homepage: [https://doby-xu.github.io/WithAnyone/](https://doby-xu.github.io/WithAnyone/)
- Online demo: [https://huggingface.co/spaces/WithAnyone/WithAnyone_Demo](https://huggingface.co/spaces/WithAnyone/WithAnyone_Demo)
---
## Wider Ecosystem: Monetize Your AI Creativity
Projects like WithAnyone integrate perfectly with platforms like [AiToEarn官网](https://aitoearn.ai/):
**AiToEarn** is an **open-source global AI content monetization platform** allowing creators to:
- Generate AI content
- Publish across multiple channels: Douyin, Kwai, WeChat, Bilibili, Rednote (Xiaohongshu), Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, X (Twitter)
- Access analytics
- See AI model rankings
Explore:
- Blog: [AiToEarn博客](https://blog.aitoearn.ai)
- AI Rankings: [AI模型排名](https://rank.aitoearn.ai)
---
## Conclusion
WithAnyone is more than a technical breakthrough — it’s AI learning to **share the frame**, producing group photos full of **warmth and story**.
From today: **Go anywhere, with anyone.**
Capture your own unique AI group photo with **WithAnyone**.
---
**References:**
[Read Original](2651001914)
[Open in WeChat](https://wechat2rss.bestblogs.dev/link-proxy/?k=87ccf96d&r=1&u=https%3A%2F%2Fmp.weixin.qq.com%2Fs%3F__biz%3DMzA3MzI4MjgzMw%3D%3D%26mid%3D2651001914%26idx%3D3%26sn%3D3992822e7056e3e258be585443966144)
