It's a int8-convrot quantization of the <a target="_blank" rel="ugc" href="https://civarchive.com/models/2930871/anteros-xxx-krea2?modelVersionId=3317072">Anteros XXX Krea2 DiT</a>, made for convenience. Enjoy!The <a target="_blank" rel="ugc" href="https://civarchive.com/models/2727641/krea-2-turbo-lora-256dim">raw-to-turbo-256-rank LoRA</a> (select the ComfyUI one, if you use that) helps a lot with the speedy generation for the Raw and Unaligned versions (reduces waiting to just 8 steps).All the credits go to the creators of this model and to the Krea2 team, for their OG monster of a DiT.Please use this model responsibly.

Anteros XXX Krea2

It seems this model really 'insists' on using "euler/simple" sampler and scheduler respectively combo. Beta or Beta57 schedulers seem to create unneeded noise in the image.Of course, it's rank 256.I added the second int_convrot DiT. It's the same-sourced bf16 model, but converted with <a target="_blank" rel="ugc" href="https://github.com/cyberdeliaAI/ComfyUI-Cyberdelia-Quantizer">Cyberdelia Quantizer</a>, whereas the first one is converted with <a target="_blank" rel="ugc" href="https://github.com/alperktt/Krea-2-SVDQuant-ComfyUI">SVDQuant-ComfyUI</a>. Please test and choose! :)I did this experiment, as I noticed the both DiTs converted with these different algorithms had also different resulting SHA256 checksums.I'll probably test and do that for the Raw and Unaligned DiTs as well.

Files with SHA256: f1f46d5e6510f04167919f166b355cc85298e2a89306a5bc4abb82ab76ae28f5