Flow Models

The aim of this repository is to test and implement Flow-Matching-based models

Warning

The data as well as dataset can not be shared!

text2img generation example

Rectflow $^1$ full steps

Rectflow $^1$ 5 steps

Rectflow $^1$ 1 step

Rectflow $^2$ 5 steps

Rectflow $^2$ 1 step

Rectflow $^2$ + Distill 1 step

Rectflow $^2$ + Distill + Midpoint 1 step

LADD 4 steps

Shortcut 4 steps

Inversion example

Input Image | FM Inversed | Changed prompt :------------------------------------------------------:|:------------------------------------------------------------:|:---------------------------------------: <img src="images/puppy.png" alt="drawing" width="350"/> | <img src="images/puppy_rec.png" alt="drawing" width="350"/> | <img src="images/kitten.png" alt="drawing" width="350"/>

TODO:

• [x] Train a foundint text2img model
• [x] FM Inversion
• [x] Midpoint, Euler, RK4 solvers (with CFG)
• [x] N-th Reflow (for straightening and fast simulation) (Same as Knowledge Distillaiton for Diffusion)
• [x] Rectified Flow Distillation (1 step)
• [x] Immiscible Diffusion
• [x] Lattent Adversarial Diffusion Distillaiton
• [x] Shortcut Model
• [ ] Consisntency Distillation
• [ ] Phased (Multiboundary) Consisntency Distillation
• [ ] Diffusion RLHF
• [ ] Bridge Matching