@eggman.eth

I hath been summoned @czar depends on your use-case! If you’ve a lot of VRAM, Flux 1.D and Flux 2 are options; they serve well for most styles including realism, and text rendering. If you’ve boatloads of VRAM, Qwen Image is the best-in-class imo. Excels in all styles, extremely good prompt adherence, and excellent text rendering to boot. Either of the above with quantisation will serve fine if you have at least 24GB of VRAM. Results may vary if you go much lower. If you’re on a VRAM budget, z-image-turbo. There’s debate on whether or not this is better than Flux (1 & 2). It’s subjective, but personally I think z-image is far more impressive. For perfect realism, use a video model like WAN2.2 and generate the first frame of a video to use as an image. Looks amazing. But RAM hungry too. If you’re on more of a VRAM budget, Chroma is a good high-speed model with excellent range, and respectable text rendering. If you want nsfw, then SDXL fine-tunes still rule the roost there. Bright side: they’re all very light on VRAM. We’ve everything above except chroma and z-image live on imgnai.com if you want to try them out (or leverage the SOTA models) before building your local setup for them! All our in-house models also use my custom gpu worker pipeline, allowing for stuff like fully native 4K on SDXL and the likes! 👀

3 replies

2 recasts

6 reactions