| license: apache-2.0 | |
| # Geometric Lookup Flow Matching (GLFM) | |
| The prototype formula for the next experiment. | |
| ``` | |
| Standard: x_t → UNet(x_t, t, c) → v_pred | |
| ↑ full 16384-dim path | |
| GLFM: x_t → encoder → project to S^15 → triangulate against anchors | |
| ↓ | |
| 768-dim geometric address (continuous, not discrete) | |
| ↓ | |
| conditioned_generator(address, t, c) → v_pred | |
| ``` | |
| # Prelim | |
| This one is going to attempt a new form of prediction based on velocity flow matching using spherical representation. | |
| The alignment for the early prototype is essentially going to be euler discreet ODE flow matching, and it will be more than enough for now. | |
| We'll work from there and build increased representation as needed. | |
| I'm going to attempt to fully replace the diffuser's substructure with a spherical representation. Standby. | |
| First proto is pretty much vpred, then I'll enhance the system with shift and so on. Simple improvements. |