2025年1月12日 · 通过基于前向传播复杂度的 Gflops 测量来分析扩散transformer（DiTs）的可扩展性。具有更高 Gflops 的 DiTs — 通过增加transformer的深度/宽度或增加输入token的数量会具有较低的 FID。

The back kick, or dwi chagi in Korean, is an essential Taekwondo move, and one of the most powerful strikes in martial arts. It is a powerful technique that can be used to surprise opponents and catch them off guard. The back kick is used to strike an opponent from behind, while they are unaware of your presence.

DiT块是模型的核心，它处理输入的token序列。 研究者们探索了四种不同的transformer块设计，以处理条件输入（如噪声时间步t、类别标签c等）。 设计包括：

Back Kick(Dwi Chagi): A back kick(also referred to as a turn sidekick) is a powerful kick in TaeKwondo. Your back leg is used in this kick and you turn 180 degrees to gain power/speed.

In preparing for your black belt grading regardless of dan level, you will need to explain the meaning of your current poomsae and learn Korean terminology used in Taekwondo. To further assist your preparation to obtain a black belt, you can download one of our handy Dan Grading Pocket Guides from the Downloads page.

Tuit Chagi (Patada Hacia Atrás): Se ejecuta girando el cuerpo y extendiendo la pierna hacia atrás, golpeando con el talón. Nako Chagi (Patada en Gancho): Consiste en levantar la pierna y enganchar con el talón, generalmente hacia la cabeza del oponente.

A chagi is a technique to overpower the opponent by making use of the foot manipulated by the motion of leg. The chagi can be performed by using the force at the time of extending the bent and folded knee or the flexion force of the extened leg, making use of the body's turning force.

The Back Kick ( 뒤차기 dwi-chagi ) is an intermediate level technique typically introduced during yellow belt (8th geup) to green belt (6th geup) levels when the taekwondo practitioner has some kicking ( 차기 chagi ) experience already.

Dolyo Chagi (dol-yo cha-gee): Turning Kick. Naeryo Chagi (na-ree-oh cha-gee): Axe Kick. Miro Chagi (mee-ro chagi): Push Kick. Yop Chagi (yop cha-gee): Side kick.

本文探索了一类新的基于 Transformer 的扩散模型 Diffusion Transformers (DiTs)。 本文训练 latent diffusion models 时，使用 Transformer 架构替换常用的 UNet 架构，且 Transformer 作用于 latent patches 上。 作者探索了 DiT 的缩放性，发现具有较高 GFLOPs 的 DiT 模型，通过增加 Transformer 宽度或者深度或者输入 token 数量，始终有更好的 FID 值。 最大的 DiT-XL/2 模型在 ImageNet 512×512 和 256×256 的测试中优于所有先前的扩散模型，实现了 2.27 的 FID 值。