Lakonik/LakonLab
Official implementation of AsymFlow, pi-Flow, GMFlow
What it solves
LakonLab provides a high-performance framework for experimenting with and training large diffusion models. It specifically addresses the need for efficient few-step image generation and editing by implementing advanced flow-matching techniques that reduce the number of sampling steps required to produce high-quality images.
How it works
The codebase implements three primary research projects:
- AsymFlow: Asymmetric flow models for image generation.
- pi-Flow: A policy-based imitation distillation method that enables few-step generation (e.g., 4-step sampling for FLUX and Qwen-Image).
- GMFlow: Gaussian Mixture Flow Matching models.
To support these, LakonLab includes performance optimizations like seamless switching between DDP, FSDP, and FSDP2, weight tying for LoRA fine-tuning to save memory, and advanced flow SDE solvers (FlowSDEScheduler and FlowMapSDEScheduler) for flexible sampling.
Who it’s for
It is designed for AI researchers and developers working with large-scale diffusion models, specifically those interested in flow matching, distillation for faster inference, and high-performance training infrastructure.
Highlights
- Few-Step Generation: Enables high-quality image generation in as few as 4 steps for models like FLUX.1 and FLUX.2.
- Infrastructure Optimizations: Supports gradient accumulation, mixed precision, and flexible distributed training (DDP/FSDP).
- Flexible I/O: Native support for loading checkpoints from HuggingFace, HTTP/HTTPS URLs, and AWS S3.
- Comprehensive Evaluation: Integrated support for standard metrics like FID, KID, IS, and CLIP similarity, as well as external benchmarks like GenEval and DPG-Bench.