Krea2 | RAW base model Turbo finished image chain

This is an AI text-to-image application based on Krea2 RAW base model and Krea2 Turbo relay sampling.

After users input a prompt, the system first uses the Krea2 RAW model for initial sampling to establish the main structure, spatial relationships, and basic composition of the image. Then, it switches to the Krea2 Turbo model for continued sampling, refining the initial latent into the final image.

Unlike the standard single-stage image generation of Krea2 Turbo, this process adopts a two-stage sampling structure. In the first stage, it does not produce a complete image but only runs the initial sampling and retains the remaining noise. In the second stage, it takes this intermediate latent, without re-adding noise, and directly uses the Turbo model to complete the final image. This approach leverages the foundational expressive capability of the RAW base model combined with the rapid convergence capability of the Turbo model, making it suitable for testing the impact of different base models, Turbo models, and sampling segmentation methods on image texture.

The entire process is automated: Krea2 RAW model loading, Turbo final model loading, Qwen3VL text encoding, 21:9 widescreen latent creation, RAW initial sampling, Turbo relay sampling, Qwen Image VAE decoding, image preview, and final saving.

The current default configuration is suitable for creating widescreen cinematic posters, dark fantasy character images, game concept art, horizontal covers, model relay testing, RAW and Turbo comparison tests, and Krea2 final chain experiments.