MiniSora Community

<!-- PROJECT SHIELDS -->

[![Contributors][contributors-shield]][contributors-url] [![Forks][forks-shield]][forks-url] [![Issues][issues-shield]][issues-url] [![MIT License][license-shield]][license-url] [![Stargazers][stars-shield]][stars-url] <br />

<!-- PROJECT LOGO --> <div align="center"> <img src="assets/logo.jpg" width="600"/> <div>&nbsp;</div> <div align="center"> </div> </div> <div align="center">

English | 简体中文

</div> <p align="center"> 👋 join us on <a href="https://cdn.vansin.top/minisora.jpg" target="_blank">WeChat</a> </p>

The MiniSora open-source community is positioned as a community-driven initiative organized spontaneously by community members. The MiniSora community aims to explore the implementation path and future development direction of Sora.

• Regular round-table discussions will be held with the Sora team and the community to explore possibilities.
• We will delve into existing technological pathways for video generation.
• Leading the replication of papers or research results related to Sora, such as DiT (MiniSora-DiT), etc.
• Conducting a comprehensive review of Sora-related technologies and their implementations, i.e., "From DDPM to Sora: A Review of Video Generation Models Based on Diffusion Models".

Hot News

• Stable Diffusion 3: MM-DiT: Scaling Rectified Flow Transformers for High-Resolution Image Synthesis
• MiniSora-DiT: Reproducing the DiT Paper with XTuner
• Introduction of MiniSora and Latest Progress in Replicating Sora

Reproduction Group of MiniSora Community

Sora Reproduction Goals of MiniSora

1. GPU-Friendly: Ideally, it should have low requirements for GPU memory size and the number of GPUs, such as being trainable and inferable with compute power like 8 A100 80G cards, 8 A6000 48G cards, or RTX4090 24G.
2. Training-Efficiency: It should achieve good results without requiring extensive training time.
3. Inference-Efficiency: When generating videos during inference, there is no need for high length or resolution; acceptable parameters include 3-10 seconds in length and 480p resolution.

MiniSora-DiT: Reproducing the DiT Paper with XTuner

https://github.com/mini-sora/minisora-DiT

Requirements

We are recruiting MiniSora Community contributors to reproduce DiT using XTuner.

We hope the community member has the following characteristics:

1. Familiarity with the OpenMMLab MMEngine mechanism.
2. Familiarity with DiT.

Background

1. The author of DiT is the same as the author of Sora.
2. XTuner has the core technology to efficiently train sequences of length 1000K.

Support

1. Computational resources: 2*A100.
2. Strong supports from XTuner core developer P佬@pppppM.

Recent round-table Discussions

Paper Interpretation of Stable Diffusion 3 paper: MM-DiT

Speaker: MMagic Core Contributors

Live Streaming Time: 03/12 20:00

Highlights: MMagic core contributors will lead us in interpreting the Stable Diffusion 3 paper, discussing the architecture details and design principles of Stable Diffusion 3.

PPT: FeiShu Link

<!-- Please scan the QR code with WeChat to book a live video session. <div align="center"> <img src="assets/SD3论文领读.png" width="100"/> <div>&nbsp;</div> <div align="center"> </div> </div> -->

Highlights from Previous Discussions

Night Talk with Sora: Video Diffusion Overview

ZhiHu Notes: A Survey on Generative Diffusion Model: An Overview of Generative Diffusion Models

Paper Reading Program

• Sora: Creating video from text

• Technical Report: Video generation models as world simulators

• Latte: Latte: Latent Diffusion Transformer for Video Generation

  • Latte Paper Interpretation (zh-CN), ZhiHu(zh-CN)

• DiT: Scalable Diffusion Models with Transformers

• Stable Cascade (ICLR 24 Paper): Würstchen: An efficient architecture for large-scale text-to-image diffusion models

• Stable Diffusion 3: MM-DiT: Scaling Rectified Flow Transformers for High-Resolution Image Synthesis

  • SD3 Paper Interpretation (zh-CN), ZhiHu(zh-CN)

• Updating...

Recruitment of Presenters

• DiT (ICCV 23 Paper)
• Stable Cascade (ICLR 24 Paper)

Related Work

• 01 Diffusion Model
• 02 Diffusion Transformer
• 03 Baseline Video Generation Models
• 04 Diffusion UNet
• 05 Video Generation
• 06 Dataset
  • 6.1 Pubclic Datasets
  • 6.2 Video Augmentation Methods
    • 6.2.1 Basic Transformations
    • 6.2.2 Feature Space
    • 6.2.3 GAN-based Augmentation
    • 6.2.4 Encoder/Decoder Based
    • 6.2.5 Simulation
• 07 Patchifying Methods
• 08 Long-context
• 09 Audio Related Resource
• 10 Consistency
• 11 Prompt Engineering
• 12 Security
• 13 World Model
• 14 Video Compression
• 15 Mamba
  • 15.1 Theoretical Foundations and Model Architecture
  • 15.2 Image Generation and Visual Applications
  • 15.3 Video Processing and Understanding
  • 15.4 Medical Image Processing
• 16 Existing high-quality resources
• 17 Efficient Training
  • 17.1 Parallelism based Approach
    • 17.1.1 Data Parallelism (DP)
    • 17.1.2 Model Parallelism (MP)
    • 17.1.3 Pipeline Parallelism (PP)
    • 17.1.4 Generalized Parallelism (GP)
    • 17.1.5 ZeRO Parallelism (ZP)
  • 17.2 Non-parallelism based Approach
    • 17.2.1 Reducing Activation Memory
    • 17.2.2 CPU-Offloading
    • 17.2.3 Memory Efficient Optimizer
  • 17.3 Novel Structure
• 18 Efficient Inference
  • 18.1 Reduce Sampling Steps
    • 18.1.1 Continuous Steps
    • 18.1.2 Fast Sampling
    • 18.1.3 Step distillation
  • 18.2 Optimizing Inference
    • 18.2.1 Low-bit Quantization
    • 18.2.2 Parallel/Sparse inference

<h3 id="diffusion-models">01 Diffusion Models</h3>
Paper	Link
1) Guided-Diffusion: Diffusion Models Beat GANs on Image Synthesis	NeurIPS 21 Paper, GitHub
2) Latent Diffusion: High-Resolution Image Synthesis with Latent Diffusion Models	CVPR 22 Paper, GitHub
3) EDM: Elucidating the Design Space of Diffusion-Based Generative Models	NeurIPS 22 Paper, GitHub
4) DDPM: Denoising Diffusion Probabilistic Models	NeurIPS 20 Paper, GitHub
5) DDIM: Denoising Diffusion Implicit Models	ICLR 21 Paper, GitHub
6) Score-Based Diffusion: Score-Based Generative Modeling through Stochastic Differential Equations	ICLR 21 Paper, GitHub, Blog
7) Stable Cascade: Würstchen: An efficient architecture for large-scale text-to-image diffusion models	ICLR 24 Paper, GitHub, Blog
8) Diffusion Models in Vision: A Survey	TPAMI 23 Paper, GitHub
9) Improved DDPM: Improved Denoising Diffusion Probabilistic Models	ICML 21 Paper, Github
10) Classifier-free diffusion guidance	NIPS 21 Paper
11) Glide: Towards photorealistic image generation and editing with text-guided diffusion models	Paper, Github
12) VQ-DDM: Global Context with Discrete Diffusion in Vector Quantised Modelling for Image Generation	CVPR 22 Paper, Github
13) Diffusion Models for Medical Anomaly Detection	Paper, Github
14) Generation of Anonymous Chest Radiographs Using Latent Diffusion Models for Training Thoracic Abnormality Classification Systems	Paper
15) DiffusionDet: Diffusion Model for Object Detection	ICCV 23 Paper, Github
16) Label-efficient semantic segmentation with diffusion models	ICLR 22 Paper, Github, Project
<h3 id="diffusion-transformer">02 Diffusion Transformer</h3>
Paper	Link
1) UViT: All are Worth Words: A ViT Backbone for Diffusion Models	CVPR 23 Paper, GitHub, ModelScope
2) DiT: Scalable Diffusion Models with Transformers	ICCV 23 Paper, GitHub, Project, ModelScope
3) SiT: Exploring Flow and Diffusion-based Generative Models with Scalable Interpolant Transformers	ArXiv 23, GitHub, ModelScope
4) FiT: Flexible Vision Transformer for Diffusion Model	ArXiv 24, GitHub
5) k-diffusion: Scalable High-Resolution Pixel-Space Image Synthesis with Hourglass Diffusion Transformers	ArXiv 24, GitHub
6) Large-DiT: Large Diffusion Transformer	GitHub
7) VisionLLaMA: A Unified LLaMA Interface for Vision Tasks	ArXiv 24, GitHub
8) Stable Diffusion 3: MM-DiT: Scaling Rectified Flow Transformers for High-Resolution Image Synthesis	Paper, Blog
9) PIXART-Σ: Weak-to-Strong Training of Diffusion Transformer for 4K Text-to-Image Generation	ArXiv 24, Project
10) PIXART-α: Fast Training of Diffusion Transformer for Photorealistic