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.
This is an extremely early experimental checkpoint. It may not have learned anything useful, may be worse than the base model, and has not been evaluated for quality or safety. Request access only if you understand these limitations.
Log in or Sign Up to review the conditions and access this model content.
Krea2 Dan2M Full Fine-Tune, 3 Epochs (Extremely Early)
Warning: this is an extremely early experimental model. It may not have learned anything useful at all. Quality, prompt adherence, concept coverage, memorization, stability, and safety have not been evaluated systematically. It may perform worse than the original Krea 2 model. Do not treat this as a finished or production-ready release.
This repository contains the raw DiT weights from a three-epoch,
full-parameter fine-tune of krea/Krea-2-Raw. It is published to make very
early testing possible, not as a claim that the training recipe or resulting
model is effective.
Files
krea2_dan2m_fullft_3ep_early.safetensors: BF16 raw DiT checkpoint.training_config.toml: sanitized training configuration and reproducibility notes. Local storage paths and dataset manifests are intentionally omitted.
The VAE, text encoder, tokenizer, and other pipeline components are not included. Use the corresponding components and inference implementation from the original Krea 2 release.
Training summary
| Parameter | Value |
|---|---|
| Base model | krea/Krea-2-Raw |
| Training type | Full-parameter DiT fine-tuning |
| Epochs | 3 |
| Optimizer steps | 16,221 |
| Final reported average loss | approximately 0.097 |
| Resolution | Bucketed, up to 1024 |
| GPUs | 8 |
| Batch size | 24 per GPU |
| Gradient accumulation | 3 |
| Effective global batch | 576 |
| Optimizer | Fused AdamW |
| Learning rate | 1e-4 |
| Scheduler | Warmup-stable-decay |
| Stored checkpoint precision | BF16 |
| Training compute | MXFP8 with BF16 mixed precision |
| Attention / compilation | Flash Attention, torch.compile |
| Seed | 20260824 |
The text-fusion attention path was kept in FP32 during training to reduce the risk of numerical overflow.
Training data notes
The run used the experimental Dan2M mixture prepared for this project, including a large Danbooru-derived component and additional curated visual sources. Captions and dataset manifests are not distributed in this repository. No claim is made that the mixture is balanced, complete, deduplicated, or free from problematic content.
Usage
This is a raw replacement DiT checkpoint, not a LoRA and not a complete Diffusers pipeline. Load it in a Krea 2-compatible inference stack in place of the base DiT while retaining the original Krea 2 VAE, text encoder, tokenizer, and sampling implementation.
No recommended sampler, guidance scale, resolution, or prompt format has been established for this checkpoint. Begin with the original Krea 2 inference settings and compare directly against the base model.
Limitations
- The checkpoint may not have learned meaningful improvements.
- It may regress anatomy, composition, text rendering, realism, style control, or prompt adherence.
- It may reproduce biases, artifacts, or memorized elements from its training distribution.
- It has not been evaluated for safety, factuality, identity preservation, or suitability for any downstream use.
- Three epochs and the reported training loss are not evidence of model quality.
Access is gated with automatic approval to make the experimental status visible before download. Approval does not imply that the model is safe, capable, or recommended.
License
Use is subject to the license and usage terms of the original Krea 2 base model.
The base model declares an other license on Hugging Face. Users are responsible
for reviewing and complying with the upstream terms and applicable law.
Integrity
The SHA256 checksum is recorded in SHA256SUMS.
Model tree for SumomoLee/Krea2-Dan2M-FullFT-3ep-Early
Base model
krea/Krea-2-Raw