Update README.md
Browse files
README.md
CHANGED
|
@@ -51,36 +51,6 @@ Our code also support LogNorm and VAR(Various Aspect Ratio Training)
|
|
| 51 |
[//]: # ()
|
| 52 |
[//]: # ()
|
| 53 |
|
| 54 |
-
## Linear-Multi-step Solvers and NeuralSolvers
|
| 55 |
-
We also provide a adams-like linear-multi-step solver for the recitified flow sampling. The related configs are named with `adam2` or `adam4`. The solver code are placed in `./src/diffusion/flow_matching/adam_sampling.py`.
|
| 56 |
-
|
| 57 |
-
Compared to Henu/RK4, the linear-multi-step solver is more stable and faster.
|
| 58 |
-
|
| 59 |
-
During some experiments, we supringly find that the linear-multi-step solver can achieve comparable results even with FlowTurbo.
|
| 60 |
-
|
| 61 |
-
As they are distinct methods, so armed with Adams, we believe FlowTurbo can be more powerful.
|
| 62 |
-
|
| 63 |
-
Also, We provide some magic solvers for the recitified flow sampling. These solvers are highly inspired by linear-multi-steps methods, and consists of just some **Magic Numbers**
|
| 64 |
-
These solvers are really powerful and interesting. We place the related code in `./src/diffusion/flow_matching/ns_sampling.py`.
|
| 65 |
-
|
| 66 |
-
| SiT-XL-R256 | Steps | NFE-CFG | Extra-Paramters | FID | IS | PR | Recall |
|
| 67 |
-
|--|-------|----------|-----------------|------|-------|------|--------|
|
| 68 |
-
| Heun | 8 | 16x2 | 0 | 3.68 | / | / | / |
|
| 69 |
-
| Heun | 11 | 22x2 | 0 | 2.79 | / | / | / |
|
| 70 |
-
| Heun | 15 | 30x2 | 0 | 2.42 | / | / | / |
|
| 71 |
-
| Adam2 | 6 | 6x2 | 0 | 6.35 | 190 | 0.75 | 0.55 |
|
| 72 |
-
| Adam2 | 8 | 8x2 | 0 | 4.16 | 212 | 0.78 | 0.56 |
|
| 73 |
-
| Adam2 | 16 | 16x2 | 0 | 2.42 | 237 | 0.80 | 0.60 |
|
| 74 |
-
| Adam4 | 16 | 16x2 | 0 | 2.27 | 243 | 0.80 | 0.60 |
|
| 75 |
-
| FlowTurbo | 6 | (7+3)x2 | 30408704(29M) | 3.93 | 223.6 | 0.79 | 0.56 |
|
| 76 |
-
| FlowTurbo | 8 | (8+2)x2 | 30408704(29M) | 3.63 | / | / | / |
|
| 77 |
-
| FlowTurbo | 10 | (12+2)x2 | 30408704(29M) | 2.69 | / | / | / |
|
| 78 |
-
| FlowTurbo | 15 | (17+3)x2 | 30408704(29M) | 2.22 | 248 | 0.81 | 0.60 |
|
| 79 |
-
| NeuralSolver | 6 | 6x2 | 21 | 3.57 | 214 | 0.77 | 0.58 |
|
| 80 |
-
| NeuralSolver | 7 | 7x2 | 28 | 2.78 | 229 | 0.79 | 0.60 |
|
| 81 |
-
| NeuralSolver | 8 | 8x2 | 36 | 2.65 | 234 | 0.79 | 0.60 |
|
| 82 |
-
| NeuralSolver | 10 | 10x2 | 55 | 2.40 | 238 | 0.79 | 0.60 |
|
| 83 |
-
| NeuralSolver | 15 | 15x2 | 110 | 2.24 | 244 | 0.80 | 0.60 |
|
| 84 |
|
| 85 |
## Citation
|
| 86 |
```bibtex
|
|
|
|
| 51 |
[//]: # ()
|
| 52 |
[//]: # ()
|
| 53 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 54 |
|
| 55 |
## Citation
|
| 56 |
```bibtex
|