| tags: [flow-matching, generative, anime, unconditional-image-generation] | |
| # Flow Matching — anime faces 64x64 | |
| Rectified/conditional flow matching. Predicts velocity `v = x1 - x0` along the | |
| straight path `xt = (1-t)*x0 + t*x1`, where `x0` is noise and `x1` is data. | |
| - **UNet**: ch=64, mults=(1,2,2,4), attn @16 — ~16.3M params | |
| - **Trained**: 30 epochs, final FM loss 0.2059 | |
| - **Sampling**: Euler ODE, `x += v(x,t)*dt`, ~10 steps is enough | |
| - Checkpoint contains `net`, `ema`, `opt`, `scaler`, `losses` — resumable. | |
| Use the `ema` weights for sampling. | |