Krea2 Biphasic Formation Technique | RAW Draft × Turbo Convergence

This is a dual-stage AI text-to-image application based on Krea2 RAW drafting and Krea2 Turbo refinement.

After the user inputs a scene prompt, the system first uses the Krea2 RAW model to generate a basic latent, establishing the subject, composition, spatial relationships, and initial visual direction. Then, the process hands this RAW latent to the Krea2 Turbo model for a second round of refinement at a lower denoise level, ultimately outputting a more stable, complete, and polished image.

Unlike the regular Krea2 Turbo single-stage image generation, this process splits the generation into two stages. The RAW stage handles drafting and exploration, while the Turbo stage focuses on refining, consolidating, and finalizing the output. This allows simultaneous testing of the RAW model's original expressive capabilities and the Turbo model's ability to enhance details, style, and completeness under low-intensity refinement.

The entire process is automated: Krea2 RAW model loading, Krea2 Turbo model loading, Qwen3VL text encoding, 9:21 ultra-vertical latent creation, RAW first-stage sampling, RAW intermediate image saving, Turbo second-stage low-intensity refinement, Qwen Image VAE decoding, final image preview, and saving.

The current default configuration is suitable for creating ultra-vertical concept portraits, fashion art posters, surreal character images, mobile wallpapers, short video thumbnails, RAW and Turbo comparison tests, Krea2 dual-model relay tests, and high-quality text-to-image experiments.