Tensor.Art
Create

LustiMix

CHECKPOINT
Original

Updated:

BigLustiMix_Rebirth
V1

Credits by OxFires

https://civitai.com/models/1352908/lustimix?modelVersionId=1528162

It's just a mix merge of the best models from the Lustify collection created by coyotte.

This merge contains the OLT + Endgame + Alpha V4 version.

Some people have complained that the OLT version looks too "synthetic", I'm not entirely agree, this version is in fact very good but requires a bit of time to understand and adjust the prompts.

But since I like experimenting and trying things out, I thought why not mix the best of Lustify and see what happens.

According to my tests, with models such as lustify / BigASP... You can push the resolution up to 2048px (Even if I prefer to stay in my safety zone around 1800px) without using HiresFix but by using Kohya HRFix. This allows you to gain a fair amount of definition and detail without encountering too many distortions (even if this happens more often) and also means faster generations.

Just try to start from this basic ratio of ‘832x1216’ and gradually work your way up while maintaining this ratio.

Why this ratio? Because nutbutter has used a lot of reference images with this resolution to train BigASP, so I may naively think that I have a better chance of getting good images if I use a higher resolution with the same ratio

I always use this useful plugin if I want to lock and keep my ratio.

I also strongly recommend changing the "inpaint width and height" settings in Adetailer to 1024 and the use of the "mediapipe_face_mesh" model. Obvious for some, perhaps less so for others

Version Detail

SDXL 1.0
About this version A 3rd merged version, which is the result of two completely experimental merges but I knew where I wanted to go so I did several tests until I got the recipe right. (I spent almost 5 days on it). I also learnt a lot about how U-net blocks were designed for SDXL. For the first merge I used BigLustiMix for model A and BigAsp2.0 for model B, using the MBW option with the ‘GRAD_A’ preset and the ‘cosine A’ calculation mode. (The result of this merge is too ‘smooth’, too washed out. so i just wanted a little ‘something extra’ to make it look better.) For the second merge I again used BigLustiMix in model A with the result of the first merge, still using MBW but this time with the ‘GRAD_V’ preset, still in cosine A. I really wanted a real in-depth mix by applying this method, still keeping the Lustify base and BigASP2.0, knowing that the last Lustify OLT model would retain its impact in this merge. I tried a 3rd merge in every way possible using the NatVision2.7 model but the results I saw didn't suit me at all so I gave up. It's easy enough to try out my merges and see whether they're stable or not: I always use Kohya HRfix to get as high a resolution as possible (always starting from the base 1216 x 832), then I use at least two LoRAs to see how it behaves. Then I check how the generation behaves, especially with two peoples. if the arms or legs are not too fused and the poses are coherent according my prompts, I validate my tests. The aim of this merge was to give it an even more photographic and analog/amateur feel, while at the same time still ensuring that it incorporated the latest improvements to Lustify (endgame + OLT) while wanting BigASP 2.0 to have a good impact on this merge. Because honestly nutbutter has done some incredible work.

Project Permissions

    Use Permissions

  • Use in TENSOR Online

  • As a online training base model on TENSOR

  • Use without crediting me

  • Share merges of this model

  • Use different permissions on merges

    Commercial Use

  • Sell generated contents

  • Use on generation services

  • Sell this model or merges

Related Posts