Baseline concept training
Contents 路 Next: Training the assistant: v1 and v2
At 500 updates, ordinary 512px training produced a recognisable Domokun in the beach scene, with unusual geometry. At 1,000 updates, both original scene prompts returned people and the unrelated fox acquired a large toothed mouth.
| Setting | Value |
|---|---|
| Training | Domokun only, 512px, 1,000 updates |
| Learning rate | 1e-4 |
| Assistant | None |
Training progression
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 |  |  |  |  |
| Update 500 |  |  |  |  |
| Update 1,000 |  |  |  |  |
Prompt and guidance comparison
With the character noun and CFG 4, the final checkpoint produces Domokun-like imagery with facial-placement and shape errors; the fox is conventional.
512px output 路 40 inference steps 路 true two-pass CFG 4 路 CUDA-generator seeds shown below 路 Details-fixed VAE (madebyollin) 路 full-frame decoding. Seed: 123.
Final checkpoint, CFG 4, seed 123
| Checkpoint / method | Beach (original prompt) | Beach | Red scarf | Fox | Portrait |
|---|---|---|---|---|---|
| Base |  |  |  |  |  |
| Unassisted 512, update 1,000 |  |  |  |  |  |
Reproduce
Training configuration 路 Datasets 路 Prompts
500 updates 路 1,000 updates 路 Shared training and evaluation settings
Limitations
Prompt wording and CFG change together in the second comparison, so their individual effects remain unresolved.