加油，争取出个衍生大模型，让老外看看国人的实力，绅士永不妥协！

This is really just using SD3 as a guide to run any SDXL model over it. All the work is being done in the refiner. This is more of a cheat than anything else. Might as well just stick with SDXL.

This is great, try to replicate this workflow, but how do I get the VAE for sd3?

Nice but, unfortunately, it doesn't fix the hands

Why is there a lot of noise in the graph I generated?

This genuinely could be the best way we can use SD3 2B right now! as just a pre noise conditioner for SDXL and not just for NSFW, it helps with hands and everything, while having the prompt adherence of SD3.

It works, as intended and "fits" in 32GB of RAM and 12GB of VRAM. Not sure how much od SD3 we have or maybe STOIQO AfroditeXL checkpoint is so good, as I haven't had a chance to test it standalone.

I added a samplercustomadvanced with nvidia align your steps sigmas and this is working really well. (the kijaj version of AYS not comfyui) though I can't use start_at_step but it still works. I know it's sdxl technically, but I'm getting much more variety out of the outputs and better prompt adherence. I just put something close to the idea on the left in SD3 prompt, and what I actually want on the right in SDXL prompt. Great idea with the latent input. I would only recommend allow us to adjust the denoise in your ksampler or implement a modified class samplercustomadvanced with ays, the extra inputs and a denoise strength option.

I'm wondering, why you use external primitive nodes for text ? It could as well be directly in the Clip text encodes, and it's not being re-used for other nodes so what are the benefits doing that ?

Anyone have a link for that SD3 VAE ? I can't find it anywhere (is that even useful?)
Please feel free to leave me a message here or in tchat if necessary, thanks!

What are vajayjays, chopped liver?