---

---

动机 Motivation

在微信之前，QQ 是中国最常用的网络通讯工具。QQ 提供了一系列系统头像，其中的许多成为了经典，甚至不少老用户至今还用着这些系统提供的头像。

回顾 QQ 头像的发展历史，大致分为 QQ2000、QQ2004、QQ2007 三个阶段。其中，QQ2007 版本迭代幅度最大，QQ 头像的品质也得到升级。这批头像一直延用到 QQ2015，伴随着我的青春期，至今看到仍然觉得亲切。因此，我尝试训练画风模型，回味复古头像的感觉。

Before WeChat, QQ was the most commonly used messaging app in China. It offered a range of avatars, many of which became classics, and even many users still use these system-provided avatars today.

Looking back at the development history of QQ avatars, it is roughly divided into three stages: QQ2000, QQ2004 and QQ2007. With the major update of QQ2007, QQ avatar quality had been upgraded dramatically. These avatars have been used until QQ2015, accompanying my adolescence, and I’m still familiar with them. Therefore, I try to train a style model to recall the memories about retro QQ avatars.

创造属于你的经典头像 Create a classic avatar of yours own

通过解包 QQ 得到 105 个 350×350 头像，并批量放大到 512×512 构成了训练数据集（感谢 (′～`●)@bilibili 搜集的数据）。使用 GPT-4o 对数据打标，并在 anything-mix SD 1.5 上进行训练。

105 avatars of 350×350 pixels were obtained by unpacking QQ and batch scaled to 512×512 to establish the training dataset (thanks to (′ ~ ` ●)@bilibili for collecting the data). The data were labeled with GPT-4o and trained on anything-mix SD 1.5.

以 IP-Adapter 为核心搭建工作流，我们可以提取自拍的特征并创建带有 QQ 头像特征的图像。为了保证生成的头像不受到自拍的背景影响，我选择先去掉背景。我还引入了 LLaVA 来自动生成 Prompt， 以更好地保留特征。

Utilizing IP-Adapter, we can craft avatars that closely resemble our selfies. To ensure that the avatar won’t generate errors due to the background of the selfie, I removed the background before inferencing. I also introduced LLaVA to generate the prompt to better preserve the features.

如果QQ允许用户创建这样的复古头像，我觉得会很有意思。

I think it would be interesting if QQ allowed users to create retro avatars like this.