Krea2 Turbo Text-to-Image Image-to-Image

Dual Sampling Prompt Optimization Full Workflow

Krea2 is a text-to-image diffusion model that generates images based on natural language descriptions. The model is designed to support applications such as creativity, business, development, and research, including image generation, concept design, design exploration, visual production workflows, and integration into applications and creative tools.

Krea2 self-developed DiT architecture, outstanding native aesthetics
12B parameter self-developed model, trained on high-quality artistic datasets, featuring advanced color tones, natural lighting and shadow effects, low error rates in character and scene structures, paired with the Qwen encoder for accurate understanding of Chinese prompts and high adherence to complex instructions.
Dual versions tailored for different creative needs
Krea2 Raw: High-quality LoRA training base, suitable for refined commercial output and style fine-tuning;
Krea2 Turbo: 8-step rapid sampling, high output efficiency, suitable for batch inspiration drafts, with interchangeable LoRA weights.
Exceptional style transfer capability
Supports precise replication of color tones, art styles, and compositions using single/multiple reference images, compatible with all styles, easily achieving visual consistency across series works, making it ideal for commercial design and IP character creation.
Lightweight deployment friendly
Offers FP8, INT8, and other quantized versions, significantly reducing VRAM usage, hardware acceleration supported on RTX40/50 series, operable with just 8G VRAM, natively compatible with ComfyUI's full plugin ecosystem, with low barriers for local deployment.
High commercial freedom
Supports multi-aspect ratios, direct output at resolutions up to 2048, adopts Apache 2.0 license, free for personal commercial use, suitable for most visual creation scenarios including posters, portraits, illustrations, concept designs, and more.