Update README.md
Browse files
README.md
CHANGED
|
@@ -15,12 +15,47 @@ This repository provides BF16 and FP8 versions of **Krea 2 OSS (Base & Turbo)**.
|
|
| 15 |
|
| 16 |
**Watch us on Youtube:** [@VantageWithAI](https://www.youtube.com/@vantagewithai)
|
| 17 |
|
| 18 |
-
##
|
| 19 |
-
- **Original Model Creators**: All credit goes to [KREA.ai](https://www.krea.ai) for the original research, architecture, and weights.
|
| 20 |
-
- **License**: This model is subject to the **KREA 2 License Agreement**. Please read and comply with the official license terms before using these weights: [KREA 2 Licensing Terms](https://www.krea.ai/krea-2-licensing).
|
| 21 |
-
- **Purpose**: This repository is a community-contributed utility. It does not claim ownership of the original model or architecture. Its sole purpose is to provide optimized, consumer-hardware-friendly weights for the open-source community.
|
| 22 |
|
| 23 |
-
-
|
| 24 |
-
|
| 25 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 15 |
|
| 16 |
**Watch us on Youtube:** [@VantageWithAI](https://www.youtube.com/@vantagewithai)
|
| 17 |
|
| 18 |
+
## Model Summary
|
|
|
|
|
|
|
|
|
|
| 19 |
|
| 20 |
+
Krea 2 is a latent-diffusion image model trained from scratch with an emphasis on aesthetics
|
| 21 |
+
and stylistic control. The architecture is a single-stream multimodal diffusion transformer.
|
| 22 |
+
|
| 23 |
+
- **Transformer**: single-stream DiT, 12.9B parameters, 28 blocks at width 6144. Grouped-query
|
| 24 |
+
attention, a learned output gate, per-head QK normalization, and a 3-axis rotary embedding.
|
| 25 |
+
A text-fusion stage inside the transformer collapses twelve text-encoder hidden-state layers
|
| 26 |
+
into one conditioning stream.
|
| 27 |
+
- **Text encoder**: [Qwen/Qwen3-VL-4B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct),
|
| 28 |
+
tapped at twelve intermediate layers (text-only conditioning).
|
| 29 |
+
- **VAE**: the Qwen-Image autoencoder (`AutoencoderKLQwenImage`, f8, 16 latent channels).
|
| 30 |
+
- **Sampler**: flow matching with a resolution-aware timestep shift.
|
| 31 |
+
|
| 32 |
+
## Recommended Settings
|
| 33 |
+
|
| 34 |
+
| Model | Steps | CFG |
|
| 35 |
+
| ------- | ----- | ----- |
|
| 36 |
+
| Base | 52 | 3 |
|
| 37 |
+
| Turbo | 8 | 0 |
|
| 38 |
+
|
| 39 |
+
Resolution | 1024 x 1024 up to 2048 x 2048
|
| 40 |
+
|
| 41 |
+
## Prompting
|
| 42 |
+
|
| 43 |
+
Natural-language prompts are recommended. Long, detailed descriptions yield the best results,
|
| 44 |
+
though strong images are produced from short prompts as well. For text rendering, the words to
|
| 45 |
+
be rendered are wrapped in quotes. An optional prompt-expansion system prompt is available in
|
| 46 |
+
the upstream [krea-2-oss](https://github.com/krea-ai) repository.
|
| 47 |
+
|
| 48 |
+
## License
|
| 49 |
+
|
| 50 |
+
The weights are released under the [Krea 2 community license](https://www.krea.ai/krea-2-licensing).
|
| 51 |
+
|
| 52 |
+
## Citation
|
| 53 |
|
| 54 |
+
```bibtex
|
| 55 |
+
@misc{krea2,
|
| 56 |
+
title = {Krea 2},
|
| 57 |
+
author = {Krea},
|
| 58 |
+
year = {2026},
|
| 59 |
+
url = {https://www.krea.ai/krea-2}
|
| 60 |
+
}
|
| 61 |
+
```
|