Unconditional Text Generation on OpenWebText (OWT) length 1024
2.9Training Duration (days)MDLM
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| MDLMNFEs=1024, Batch size=16, Hardware=single NVIDIA A100 GPU, Training iterations=1M, Sequence length=10242026.05 | 2.9 | 529 | |
| FLMNFEs=1024, Batch size=16, Hardware=single NVIDIA A100 GPU, Training iterations=1M, Sequence length=10242026.05 | 3 | 97 | |
| Duo++ + Ψ-samplerNFEs=1024, Batch size=16, Hardware=single NVIDIA A100 GPU, Training iterations=1M, Sequence length=10242026.05 | 3.2 | 663 | |
| LDLMNFEs=1024, Batch size=16, Hardware=single NVIDIA A100 GPU, Training iterations=1M, Sequence length=10242026.05 | 5.1 | 40 | |
| DuoNFEs=1024, Batch size=16, Hardware=single NVIDIA A100 GPU, Training iterations=1M, Sequence length=10242026.05 | 5.5 | 537 | |
| CANDINFEs=1024, Batch size=16, Hardware=single NVIDIA A100 GPU, Training iterations=1M, Sequence length=10242026.05 | 5.6 | 142 |