Looped Diffusion Transformer:采样时动态调整循环深
Looped Diffusion Transformer: the same blocks run several times inside each denoising step, and loop depth can be changed at sampling time [R]
这篇论文展示了通过架构创新(循环深度)而非堆参数量提升扩散模型效率的新思路,且支持采样时动态调整,对追求性价比的开发者很有参考价值。
Paper: Looped Diffusion Transformer (arXiv 2609.40305, 30 Sep). Instead of making a text-to-image DiT bigger, it runs a shared group of transformer blocks several times inside each denoising step. The authors report that naive looping doesn't reliably help, and fix it with deep supervision on every loop plus self-modulating attention. Their headline claims: a 260M looped model beats a model 6.5x larger on several benchmarks with 4.9x less inference compute, and at a fixed inference budget, more loop depth helps more than more denoising steps.
论文:循环扩散 Transformer(arXiv 2609.40305,9月30日)。与其将文本到图像的 DiT 模型做得更大,它在每个去噪步骤内部多次运行一组共享的 transformer 块。作者指出,简单的循环并不能可靠地带来提升,并通过在每个循环上添加深度监督以及自调制注意力机制解决了这一问题。他们的核心主张是:一个拥有 2.6 亿参数的循环模型在多个基准测试中击败了参数量大 6.5 倍的模型,且推理计算量减少了 4.9 倍;在固定的推理预算下,增加循环深度比增加去噪步数带来的收益更大。
The B/16 code and weights are MIT (the weights are on Hugging Face under sensenova), so I ran it on an M5 Pro: PyTorch MPS in fp32, 100 Euler steps, CFG 6.0, and one seed per prompt reused at loop depths 1 to 4. It was trained at depth 4, and the README says other depths work without retraining. The grid shows five of the prompts I tried.
B/16 的代码和权重采用 MIT 许可证(权重托管在 Hugging Face 上的 sensenova 账号下),因此我在 M5 Pro 上进行了测试:使用 PyTorch MPS 后端,精度为 fp32,100 个 Euler 步,CFG 设为 6.0,每个提示词使用一个种子,并在 1 到 4 的循环深度下复用。该模型是在深度 4 下训练的,README 说明其他深度无需重新训练即可使用。网格图展示了我尝试的五个提示词的结果。
Most of the change happens before depth 3, and depths 3 and 4 come out very close on every prompt shown. The interesting part to me is that depth is picked at sampling time, the way the iteration count is in looped and recurrent-depth language models, so you can trade compute for refinement per image. On this machine an image took about 23 to 24 seconds at depth 1 and 43 to 47 at depth 4.
大部分变化发生在深度 3 之前,而在展示的每个提示词中,深度 3 和深度 4 的结果非常接近。对我来说有趣的部分在于,深度是在采样时选择的,类似于循环和递归深度语言模型中的迭代次数选择方式,因此你可以用计算量换取每张图像的细化程度。在这台机器上,深度 1 时生成一张图像大约需要 23 到 24 秒,而深度 4 时则需要 43 到 47 秒。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力