Krea's Open Krea 2 Raw Spawns 1M-Download Community LoRA in Weeks
A community-trained LoRA for Krea 2 focused on Asian character generation has crossed one million downloads, pulling the open-weights base model into the spotlight.
- The fidel1234/Krea2 LoRA, an Asian character adapter for Krea 2 Raw, has crossed 1 million downloads.
- Built with AI-Toolkit on Krea 2 Raw, MIT licensed, compatible with Raw and Turbo checkpoints in ComfyUI.
- Krea 2 is a 12B parameter single-stream Diffusion Transformer, Krea's first from-scratch foundation image model.
- Uses Qwen 3 VL text encoder, Grouped-Query Attention with sigmoid gating, SwiGLU, and 3D axial RoPE.
- Turbo variant generates 2K images in roughly 2 seconds; ranks top 10 on Artificial Analysis leaderboard.
- Recommended LoRA weight 0.8 to 1.0, 20 to 30 inference steps, with workflows embedded in sample image metadata.
At publication time, the community-built Krea2 adapter had passed one million downloads on Hugging Face. The LoRA applies a compact set of fine-tuned weights to Krea 2 Raw and specializes its output toward what the author describes as Japanese and Asian character generation. Its traction points to strong demand for narrow visual styles built on Krea’s open-weight model.
LoRA, short for low-rank adaptation, stores a small adjustment to a larger model instead of distributing another full checkpoint. Users still need the Krea 2 base weights. The adapter carries an MIT license, and the base model’s separate license continues to govern use of its weights.
Hugging Face’s download counter measures file retrievals, which can include automated systems and repeated downloads. Unique-user totals, active installations, and production deployments remain unknown, so the figure is best read as evidence of repository activity.
Krea splits speed from customization
Krea released Krea 2 on May 12, 2026, as its first image foundation model trained from scratch. On June 22, the company published the open-weight Krea-2-Turbo and Krea-2-Raw checkpoints. Krea describes both as 12-billion-parameter Diffusion Transformer models capable of producing 2K images, with Turbo optimized for generation in roughly two seconds.
| Checkpoint | Training state | Primary use | Published characteristic |
|---|---|---|---|
| Krea-2-Turbo | Post-trained and distilled | Interactive tools and latency-sensitive products | Roughly two-second 2K generation under Krea’s test conditions |
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
