Krea2 Dan2M Full-Rank LoKr
Experimental test checkpoint. This model is currently provided only for early testing. Its quality, stability, concept coverage, and recommended inference strength have not been systematically evaluated. It should not be treated as a finished release, and future revisions may change the training recipe or compatibility.
This repository contains a large-capacity LoKr adapter trained from the original Krea2 base as an experiment in broad concept learning.
Selected training parameters
| Parameter | Value |
|---|---|
| Adapter | Full-matrix LoKr |
| LoKr factor | 6 |
| Network dimension / alpha | 3072 / 3072 |
| Epochs | 1 |
| Optimizer | CAME |
| Learning rate | 3e-5 |
| Scheduler | Warmup-stable-decay; 3% warmup, final 20% decay, minimum LR ratio 0.1 |
| Precision | bf16 |
| Maximum training resolution | 1024 |
| GPUs | 8 |
| Batch size | 8 per GPU |
| Gradient accumulation | 2 |
| Effective global batch size | 128 |
| Attention / compilation | Flash Attention, torch.compile |
| Text encoder | Online encoding with an in-memory cache |
No inference strength is recommended yet. Test multiple conservative strengths and evaluate outputs for the intended prompt distribution.