ZImage Turbo used with Qwen_3_4b for CLIP. Qwen_3_4b is a language model that understands things, so this combo can do complex prompts (20k tokens?) and compose images..with okay-ish adherence (check the 1st two images) but of course it'll run for much longer then. The only drawback is that if you describe any features of an object or person which aren't supposed to be seen from the viewing angle you specify, the model WILL change the angle to make them seen. I think if you want to use a specific character with ease, you have to make a lora - because editing detailed prompts to exclude details that momentarily cannot be seen is a pain. It runs okay on a 4070, or even 3060, with decent generating times. (1280x1024 in ~30 secs). If you want previews in okay resolution, can go down to 12 seconds or even less if low. The fp8 model is just 6gb. Not censored but has shit data on nsfw items so doesn't draw them well. Loras for that are out. It does render fatties out of the box decently up to a respectable size but struggles with ssbbw sizes as it doesn't really know how to do belly aprons and fat folds and so on. Stretchmarks etc no dice either, but seems to do okay on cellulite etc. There's few loras for it yet... just released last week.