@Uzbeliss I can use it with 12 GB of VRAM without any problems, maybe not at the highest settings and quality, but it's easy to do. If you have hardware with at least 12 GB, you could test it. The problem would be more training, as I understand that you really need a lot of VRAM for WAN, much more than 12 GB. It depends on how powerful your equipment is, but if you have around 12 GB of VRAM and an RTX 3000 card or higher, there are a lot of optimizations to speed up generation with WAN. I make six-second videos in about two minutes with fusion x 14b (4070ti)

@Uzbeliss Try looking for some guides on how to use Fusion X and Sage Attention, it's an optimized model that allows you to generate much faster than normal models. Furthermore, huge steps forward have been made; until recently, generating was much more demanding. With 12 GB of VRAM and the right optimizations, you'll have fun now, with some sacrifices, but you'll still have fun!

It allows you to generate excellent videos with just a few steps, from 6 and up. If you keep the video length to a moderate one, you can achieve between 6 and 12 steps in 480x832p in just a few minutes. The processing time depends on three factors: the steps, the resolution, and the length of the video. Keep the fps low (the beauty of fusion x and sage attention is that it allows you to generate beautiful things even with 6 steps and CFG 1), don't make videos that are too long, and don't exceed 12 steps. It really takes two or three minutes to generate. Then, if you want the videos at a higher resolution, you can always upscale them later, so as not to lengthen the processing times.
Don't use load diffusion and load clip beyond FP16. If you use FP16, use the accumulation executable, which is faster, although it loses a bit of quality (but it's still nice).

@Uzbeliss Z-Image is a new model for image generation, as fast as SDXL and generates images with the quality of Flux FP8. It's literally becoming a reference model, I've fallen in love with it myself. Try looking on this site to see how many LoRa models are being released—one comes out every 30 seconds! By the way, it's NSFW by default.

@Uzbeliss You have to get used to it a bit at first, but then you'll fall in love with it. Try generating different things and using complex prompts. Even though it's a 6B, it allows for really complex prompts; for some things, it's revolutionary. It takes a while to understand, and the turbo version isn't the basic version, which has yet to be released. However, in terms of generation speed combined with complex prompts understood almost without fail, it's a model that's driving a lot of people crazy. Of course, with SDXL, you're used to years of training and thousands and thousands of LoRa and Checkpoints. ZIT has only been around for a week, so it takes time and community development to expand its capabilities. You just have to give it a chance and not stop at the first few generations, because you need to understand it.