You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Krea 2 r32 long-caption LoRA checkpoints

Private archival repository for representative checkpoints from an experimental Krea 2 LoRA run.

Training status

Training stopped after a non-finite loss at global step 21,179 when the automatic restart budget was exhausted. The planned target was 42,290 steps, so these are intermediate checkpoints rather than a finished release. Step 21,000 is the latest safe checkpoint. The selected checkpoints were checked for safetensors readability and finite tensor values, but were not ranked by visual evaluation.

The learning-rate resume fix was active for these checkpoints: weight-only resumes position the scheduler at global_step * world_size, avoiding a second warmup.

Configuration summary

Base architecture: Krea 2 RAW DiT
Network: networks.lora_krea2
Rank / alpha: 32 / 32
Precision: BF16
Optimizer: CAME
Learning rate: 1e-4
Scheduler: warmup-stable-decay, 10% warmup and 10% decay
Training resolution range: 768-1536, multi-resolution
Saved tensors per checkpoint: 792
Parameters stored per checkpoint: 117,306,024

Checkpoints

File	Global step	Role
`checkpoints/krea2-r32-step-010000.safetensors`	10,000	Early post-resume-fix milestone
`checkpoints/krea2-r32-step-015000.safetensors`	15,000	Mid-run milestone
`checkpoints/krea2-r32-step-020000.safetensors`	20,000	Late-run milestone
`checkpoints/krea2-r32-step-021000.safetensors`	21,000	Latest safe checkpoint

See checksums.sha256 for SHA-256 hashes and training-summary.json for machine-readable details.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support