Before maternity leave, I tried various things and dabbled in image generation.
Initially, I set up open webui and configured comfyUi and z-image turbo for image generation.
Someone suggested trying image generation as well, saying that conversational AI was getting stale. However, Z-image lacked detail compared to the writing prompts.
It would roughly sketch outlines, but there was a sense of incongruity and awkwardness, making me feel like an AI newbie.
So, I tried FLUX2 and KREA2 for about two weeks, both offering free trials to some extent.
As a result, I realized that the crucial elements in t2i (Text To Image) image generation are the text encoder and LoRa (LoRa).
Especially, Qwen was overwhelmingly superior in terms of the text encoder. While flux2 dev uses Mistral Small 3, KREA2 utilizes Qwen3-VL-4B, making a difference. Particularly for Korean expressions other than English, krea2 performed the best.
Generating images with approximately 1,000 characters using the same prompt yielded image quality similar to expectations. In terms of quality, there didn't seem to be a significant difference when applying various LoRAs.
Personally, I want to support KREA2 Turbo for image generation.
The reason is that it's faster than flux2 based on v100 32gb. Of course, the flux2 dev model is heavier, but considering the results and details, krea2 had more natural colors.
They say that the desired artistic style changes depending on the LoRA used for image generation. I don't have enough motivation to explore that yet, but seeing the direction of open-source image generation, I believe we'll be able to create images indistinguishable from reality.
Ultimately, I consider Qwen's consecutive victories. I wonder if there will be another open-source model with such a high cost-performance ratio in coding, image generation, and computation.
I particularly recommend KREA2 for cartoon, fantasy, and sci-fi elements. Of course, real-life images are also good.
