experiments/assistant-regularisation.md

Combining an assistant with regularisation

Contents · Previous: Synthetic regularisation · Next: Multi-scale training and REPA

Combining assistant v2 with synthetic regularisation produces coherent scenes, portraits, and fox illustrations in the inspected final samples. At both output sizes, the 512px-trained adapter expresses the character in the beach, scarf, and bicycle scenes. Body proportions and bicycle interactions remain imperfect.

Setting	Value
Training resolution	Separate 512px and 1024px runs
Assistant	Frozen v2, strength 1; disabled for validation
Regularisation target	Base + frozen assistant v2
Student	Base + frozen assistant v2 + trainable concept
Data	27 Domokun images + 834 synthetic images per resolution
Learning rate	1e-4
Updates at midpoint	439 Domokun + 561 regularisation
Updates at final	904 Domokun + 1,096 regularisation

The runs share initial concept weights and the backend-choice sequence. GPU audits verified the assisted target and restoration of the frozen-assistant training stack.

Final checkpoints

The 1024px-trained bicycle sample at 1024px output resembles a human in a creature costume. Its scarf colour also changes with output resolution.

CPU-generator seed 42 · 40 inference steps · CFG 1 · Details-fixed VAE (madebyollin) · full-frame decoding. Concept adapter only. Columns show prompts; rows show checkpoints or methods. Click a sample to view it at full resolution.

512px output

Checkpoint / method	Beach	Red scarf	Bicycle	Fox	Portrait
Base
512px training
1024px training

1024px output

Checkpoint / method	Beach	Red scarf	Bicycle	Fox	Portrait
Base
512px training
1024px training

Training progression

The following columns combine training resolution and prompt; every sample is 1024px output.

1024px output

Checkpoint / method	512px training: beach	512px training: bicycle	1024px training: beach	1024px training: bicycle
Base
Update 1,000
Update 2,000

Reproduce

512px: Training configuration · Datasets · Prompts · 1,000 updates · 2,000 updates

1024px: Training configuration · Datasets · Prompts · 1,000 updates · 2,000 updates

These recipes require assistant-aware parent capture. Use the recorded implementation and launch instructions and training provenance.

Limitations

The synthetic pools differ between training resolutions. Multi-scale and REPA experiments use a bare-base regularisation target, a different training setup.

Contents · Previous: Synthetic regularisation · Next: Multi-scale training and REPA