Krea2 A native — LR 0.0001
Next-scene LoRA experiment initialized from scratch, targeting5,000 steps with6,138 cached scene pairs/captions. Base checkpoint krea2_raw_fp8_scaled.safetensors, scaledFP8 storage/BF16 compute, rank64, real batch2, accumulation1, LR0.0001, 512-pixel buckets.
Current status
The long run was stopped safely atstep1949 for a conditioning audit. Full state1949 and the latest exported adapter1750 are preserved. It has not been restarted after these corrections.
Padding is now masked before text-fusion refiners; timestep projections are shared by token span following the stock contract. CPU regression/parity tests passed. Two correctedFP8 smokes (without/with frozenTurbo) passed10steps plus resume to12 and stockTurbo inference.
Strict gradient parity on real weights has not passed; the full-resolutionFP32 diagnostic exceeded32GB VRAM. See audits/corrected_native_20261001/RELATORIO_CORRECOES.md. Smokes establish execution, not learned quality or preserved eight-step distillation. Turbo integration remains experimental.
Checkpoint folders contain task adapters and optimizer/resume states. Artifact folders include configs, smoke reports and samples. Long-run evaluation used four held-outTurbo cases every250steps. Public backups are verified by file size and SHA256 before local pruning.
Galeria de imagens
Ver as 44 imagens e os grids de todos os testes atuais. Inclui LR0,0004, LR0,0001 até1750 e os dois smokes corrigidos no passo10.