experiments/multiscale-repa.md

Multi-scale training and REPA

Contents · Previous: Combining an assistant with regularisation · Next: Learning-rate experiments

The fixed-shift multi-scale REPA run produces coherent Domokun beach and bicycle samples with fine texture at 1024px. The comparisons cover single-resolution REPA, multi-scale training, and fixed versus automatic flow shift.

Shared setting	Value
Assistant	Frozen v2; disabled for validation
Regularisation target	Bare base, both adapters disabled
Rank / alpha / precision	32 / 32 / BF16
Optimizer / learning rate	AdamW BF16 / 1e-4
Batch / warmup / norm clipping	1 / 25 updates / 1.0
Gradient checkpointing	Interval 2
Multi-scale sampling	Domokun-512, Domokun-1024, synthetic-512, synthetic-1024: 0.25 each
REPA	DINOv2-G patch features, size 518, block 8, weight 0.5, spatial alignment, temporal distance 0
Flow shift	Fixed 3, or auto (approximately 1.7138 at 512px and 2.0008 at 1024px)

The three original multi-scale runs share initial concept weights and batch counts: 1,005 Domokun-512, 1,014 Domokun-1024, 1,036 synthetic-512, and 945 synthetic-1024. Training uses square buckets. At 4,000 updates, exposure per resolution approximately matches the 2,000-update single-resolution runs.

Final checkpoints

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 control, 2,000
512px + REPA, 2,000

1024px output

Checkpoint / method	Beach	Red scarf	Bicycle	Fox	Portrait
Base
512px control, 2,000
512px + REPA, 2,000

Multi-scale, 512px output

Checkpoint / method	Beach	Red scarf	Bicycle	Fox	Portrait
Base
Fixed shift, 4,000
Auto shift, 4,000
REPA, fixed shift, 4,000
REPA + auto shift, 4,000

Multi-scale, 1024px output

Checkpoint / method	Beach	Red scarf	Bicycle	Fox	Portrait
Base
Fixed shift, 4,000
Auto shift, 4,000
REPA, fixed shift, 4,000
REPA + auto shift, 4,000

Learning progression

Single-resolution REPA, 512px training:

512px output

Checkpoint / method	Beach	Red scarf	Bicycle	Fox	Portrait
Base
Step 1,000
Step 2,000

1024px output

Checkpoint / method	Beach	Red scarf	Bicycle	Fox	Portrait
Base
Step 1,000
Step 2,000

Multi-scale REPA, fixed shift:

512px output

Checkpoint / method	Beach	Red scarf	Bicycle	Fox	Portrait
Base
Step 2,000
Step 4,000

1024px output

Checkpoint / method	Beach	Red scarf	Bicycle	Fox	Portrait
Base
Step 2,000
Step 4,000

Recipes and checkpoints

Variant	Training resolution	REPA	Flow shift	Selected LoRA checkpoints
512px control	512	No	3	2,000
512px + REPA	512	Yes	3	1,000, 2,000
Multi-scale, fixed shift	512 + 1024	No	3	2,000, 4,000
Multi-scale, auto shift	512 + 1024	No	Auto	2,000, 4,000
Multi-scale + REPA, fixed shift	512 + 1024	Yes	3	2,000, 4,000
Multi-scale + REPA + auto shift	512 + 1024	Yes	Auto	4,000, 10,000 continuation

The Domokun run report contains the complete REPA + auto-shift trajectory through 10,000 updates. Shared training and evaluation settings.

Limitations

Each variant uses one training seed and these grids use one validation seed. The separate assistant-plus-regularisation experiment uses an assisted target. Keep that distinction when comparing methods.

Contents · Previous: Combining an assistant with regularisation · Next: Learning-rate experiments