Learning-rate experiments
Contents 路 Previous: Multi-scale training and REPA 路 Next: Longer training: regression and recovery
The assisted 512px run at 3e-4 severely degraded all four validation prompts by update 500. The unrelated fox and portrait became Domokun-like imagery.
| Setting | Value |
|---|---|
| Data / training resolution | Domokun only / 512px |
| Assistant | Frozen v1; disabled for validation |
| Learning rate / norm clipping | 3e-4 / 1.0 |
| Checkpoint | 500 updates |
CPU-generator seed 42 路 40 inference steps 路 CFG 1 路 Details-fixed VAE (madebyollin) 路 full-frame decoding. Output: 512px. Columns show prompts; rows show checkpoints or methods. Click a sample to view it at full resolution.
512px output
| Checkpoint / method | Beach (original prompt) | Scarf (original prompt) | Fox | Portrait |
|---|---|---|---|---|
| Base |  |  |  |  |
| Concept, 500 updates |  |  |  |  |
Reproduce
Training configuration 路 Datasets 路 Prompts 路 500 updates 路 Shared training and evaluation settings
Limitations
This separate run illustrates degradation at the higher learning rate; the available comparison does not isolate learning rate as its cause.
Contents 路 Previous: Multi-scale training and REPA 路 Next: Longer training: regression and recovery