Skip to content

OpenDiT

A high-performance DiT implementation for training and inference

Visit official site

What OpenDiT is

The framework is described as an open-source high-performance Diffusion Transformer implementation based on Colossal-AI, designed to improve training and inference efficiency for DiT applications. The material names text-to-video generation and text-to-image generation as the target applications.

What you can do with it

  • Run fast DiT training and inference
  • Use the complete text-to-image generation pipeline
  • Use the complete text-to-video generation pipeline
  • Combine hybrid parallelism with the FastSeq method

Performance notes

  • The project states up to 80% speedup on GPUs
  • The project states a 50% reduction in memory use
  • Single-node sequence parallelism is stated to save up to 48% of communication cost
  • Sharding the EMA model lowers memory cost further

Parallelism and kernels

  • FlashAttention, Fused AdaLN and Fused Layernorm are listed as kernel optimisations
  • ZeRO, Gemini and DDP are combined in a hybrid parallel approach
  • FastSeq is a sequence parallelism method for large activations and small parameters
  • The implementation is built on Colossal-AI

Pros & cons

✓ What we like

  • Both text-to-image and text-to-video pipelines are listed
  • Several memory saving measures are documented

! What to watch out for

  • The stated speedups are tied to GPU environments
  • FastSeq targets workloads with large activations and small parameters

FAQ

What is FastSeq?

The material describes it as a sequence parallelism method suited to DiT workloads where activations are large and parameters are small.

What speedups are stated?

The project states up to 80% speedup on GPUs, 50% memory reduction and up to 48% communication cost saving on a single node.

Last reviewed: 2026-11-15

More AI image generation tools

View all →

How we review