Generative media compute

英文原文文件：README.md

设备/设置

用于高维图像/视频/音频生成的 GPU/TPU 训练设置，采样成本和内存密集的 U-Net/Transformer 尤为突出。

瓶颈

稳定生成训练、高分辨率合成、潜空间效率、扩散采样步数和多模态数据吞吐。

适配的方法

VAE、GAN、DCGAN/pix2pix/CycleGAN/StyleGAN、DDPM/SDE 扩散、潜扩散、DALL-E 式文生图、改进 DDPM 和 DiT 将生成任务适配到可用加速器预算。

变得过时或不再中心的方法

无压缩的像素空间生成，以及纯对抗管线，在扩散/潜空间/Transformer 设置更易扩展后不再占据中心地位。

代表性论文

排名	年份	论文	优先级	状态
75	2020	Denoising Diffusion Probabilistic Models	5	downloaded / read_complete
76	2021	High-Resolution Image Synthesis with Latent Diffusion Models	5	downloaded / read_complete
77	2013	Auto-Encoding Variational Bayes	4	downloaded / read_complete
78	2014	Generative Adversarial Nets	4	downloaded / read_complete
79	2018	A Style-Based Generator Architecture for Generative Adversarial Networks	4	downloaded / read_complete
80	2020	Score-Based Generative Modeling through Stochastic Differential Equations	4	downloaded / read_complete
81	2021	Zero-Shot Text-to-Image Generation	4	downloaded / read_complete
82	2022	Scalable Diffusion Models with Transformers	4	downloaded / read_complete
83	2015	Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks	3	downloaded / read_complete
84	2016	Image-to-Image Translation with Conditional Adversarial Networks	3	downloaded / read_complete
85	2017	Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks	3	downloaded / read_complete
86	2021	Improved Denoising Diffusion Probabilistic Models	3	downloaded / read_complete
132	2026	Qwen3.5-Omni Technical Report	4	downloaded / read_complete

开放问题

• 区分媒体模型中的训练计算改进与推理/采样计算改进。