# hustvl/LightningDiT

[CVPR 2025 Oral] Reconstruction vs. Generation: Taming Optimization Dilemma in Latent Diffusion Models

Repository: https://github.com/hustvl/LightningDiT
Canonical: https://ross.abutalabs.com/products/lightningdit
Language: Python
License: MIT
License Family: permissive
Last push: 2025-12-16T06:42:41+00:00

## Health v2 (maintenance only)
Score: 47/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 57, release rhythm 35, longevity 43
- inputs: {"age_days": 609, "days_push": 260, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1529, forks 56 (observed 2026-08-28T04:04:58.747202+00:00)

## What it is
LightningDiT is a research codebase for latent diffusion models implementing VA-VAE and LightningDiT, achieving FID 1.35 on ImageNet-256 with 21.8x faster training than the original DiT. It accompanies a CVPR 2025 Oral paper on taming the reconstruction-vs-generation optimization dilemma in visual tokenizers.

## Use cases
- train a latent diffusion transformer on ImageNet-256
- reproduce state-of-the-art image generation FID results
- train a VA-VAE visual tokenizer
- make diffusion transformer research affordable on few GPUs
- experiment with visual tokenizer latent dimensions
- generate images with pretrained diffusion transformer weights

## When to choose
- you need fast, reproducible DiT-style image generation training
- you are researching latent diffusion tokenizers and the reconstruction-generation tradeoff
- you have limited GPU budget but want competitive ImageNet generation results

## When to avoid
- you need a production-ready image generation API or product
- you want text-to-image generation on arbitrary prompts rather than class-conditional ImageNet
- you are not working with PyTorch GPU research workflows

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, deep-learning, image-processing
- domain: deep-learning, machine-learning, image-processing, computer-vision
- platform: python
- tags: latent-diffusion, diffusion-transformer, vae, image-generation, research-code, cvpr-2025, gpu, linux

## Member repositories
- hustvl/LightningDiT (main) score 47

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:58.747202+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T04:31:28.230202+00:00, confidence not recorded.
  - readme: https://github.com/hustvl/LightningDiT (fetched 2026-08-28T04:04:58.747202+00:00, sha 558bd2d1139a)
- Data as of 2026-08-30T08:39:29.467469+00:00.