>>3027273Funny you should ask now. The chinks just released Z-Image Turbo today, a new 6B parameter model that's fast, small, and competes with Qwen Image, Flux, Chroma and a lot of other way bigger models.
>https://comfyanonymous.github.io/ComfyUI_examples/z_image/It seems mostly tuned for realism but can do anime too. Once we get the base model that people can train, it could finally dethrone SDXL. Pic related took 10 seconds on my 3090.
Flux 2 was released just before it, but it's bigger and slower than Flux 1, also very censored. Z-Image seems much more promising right now and is showing the power of a small model trained with a high quality dataset.
