- FLUX.1-dev has overshadowed Stable Diffusion, but ChatGPT's free image generation, especially the 'Studio Ghibli' style, became the new benchmark for AI images.
- Google's Gemini 2.5 Flash Image (codename Nano Banana) is an autoregressive model with strong prompt adherence, allowing for detailed edits and complex instructions.
- Nano Banana excels at following nuanced prompts, including JSON and HTML inputs, due to its robust text encoder and large context window (32,768 tokens).
- Despite its strengths, Nano Banana struggles with style transfer (e.g., Studio Ghibli) and has few IP restrictions, raising legal concerns.
- The model is relatively lenient with NSFW content, though it has moderation for both prompts and generated images.
- Effective prompt engineering with ALL CAPS and threats can improve adherence, but the model's performance varies across different tasks.
- Jupyter Notebooks are released to demonstrate reproducibility of the image generations described in the blog post.