Krea 2 Image Model Release and FFASR Benchmark Launch

모델 출시 | Thu Jun 25 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 2 sources

Krea released a 12B open-weights image generation model, while Hugging Face and Treble launched a far-field ASR benchmark.

Analysis

[Krea] released Krea 2, a 12B open-weights image generation model ^[1]

SOTA open-weights model at 12B parameters
Based on diffusion transformer (DiT) architecture
Foundation model optimized for creative exploration

[Krea 2] applied a multi-stage training pipeline and architectural improvements ^[1]

Stages of pretraining
midtraining
SFT
preference optimization
and RL
Integration of iREPA
improved VAE
and Qwen3-VL
Introduction of grouped-query attention (GQA) and sigmoid-gated attention
Lightweight timestep modulation and multilayer feature aggregation

[Krea 2] introduced a prompt expander and style-reference system ^[1]

Expands short
ambiguous user prompts into rich visual descriptions
Steerable from both text and image inputs
Narrows the gap between training conditioning and user expressions at inference

[Hugging Face · Treble Technologies] launched the FFASR Leaderboard ^[2]

The first open far-field ASR benchmark
Community-driven evaluation based on 14 simulated rooms
Real-world measurements and validated sim-to-real methodology
Pareto front visualization of WER and RTFx

[FFASR Benchmark] confirmed a large performance gap compared to near-field ^[2]

Far-field WER is several times higher than near-field in low-SNR environments
Reflects real-world use cases such as AI voice agents
conference room transcription
and in-vehicle assistants
Roadmap includes support for multi-talker
microphone arrays
and echo cancellation

Sources