So, its like nothing happened:) Not a big deal, you can adjut these parameters. Bigger problem, its still hard to prompt coherent sex scenes. Booru tags dont work, sentences, like fucked by man, man fucking girls, penis in xy are working, but with too much error.

Flux could be good but it's so expensive to train!

base flux is ok for what it is and now fast too with the nunchaku svd quant thing - bit of a hassle to install and dial in tho. (and the fact that you need a special lora loader for it makes selecting the 10 million flux loras available a huegpain). still, when the clothes come off, flux isn't the right tool (imo). chroma shows a lot of potential for certain content but we'll have to see where it ends up eventually. tl;dr, hello sdxl. : D

@WhyNaNWhenNyan so demanding to run too ;_;

AI literally can't not use the "it's not X, it's Y" (or any number of variations) framing device. I have so many instructions/rules and memories saved telling it to not do it, and it almost always slips through. 🤪

myfriendshawn that's probably cause there's no "negatives" equivalent for text generation like there is for images, so if you prompt it not to do something it just gets added into the context lol

Look at any picture you like that is posted and read the stats. Familiarize yourself with the terms, and also look out for upscales and upscalers.

Why? Use the LORA 'dmd2_sdxl_4step_lora_fp16', which you can download from this link:

https://civitai.com/models/1608870?modelVersionId=1820946

It's used like any other LORA. You set the CFG value to 1 and the steps to 8.

Sampling method: LCM

Schedule type: Exponential

And enjoy!

MirandoPaCuenca Oh wow I didn't know this existed. Thank you.

Only model, which is good with dmd is Biglove, other ones are extremely undetailed with it.

@antionesun101 Not in my experience, but ok. I'm very satisfied, maybe I'm not as demanding like you in my generations.

@MirandoPaCuenca I generate detailed clothes with different materials, thats where dmd lacks. There is just not enough detail, or its unfinished(because of low cfg and steps).

@Cyberdelia No. You are right. Dmd works with 8-12 steps and cfg1-1.3 the best.

Positive: fast, more realistic

Negative: lacks of details, less adherence

@Cyberdelia No, you're right. Obviously, you can't achieve the same level of detail with 8 steps and a CFG of 1 as with 40 steps and a CFG of 6 or 7. But in my case, with a 3GB VRAM GPU, waiting several minutes for each image seems like an eternity. That's why I decided to use DMD2 since I found out about it. As I said above, I'm not that particular about detail. Just being able to generate images with my GPU, I already consider myself lucky hahahaha.

@antionesun101 Simply not true. BL is great but there's a lot of dmd2 checkpoints with excellent iq.

@Cyberdelia I often use DMD2, it removes certain details, but sometimes it seems that SDXL would be better if it didn't add these details, especially when it comes to landscapes. SDXL often adds illogical or buggy details. The pictures with DMD2 seem simpler, but there are fewer errors. Also, DMD2 gives good skin detail. And in general, small details seem sharpnes

@Gaydevai_paulo DMD2 gives better anatomy and the outputs are more "photographically" realistic. On the bad side, if you convert a pony checkpoint to dmd2 it will lose a lot of the cartoon aesthetics. DMD2 is best suited for photorealistic checkpoints. The only serious con of dmd2 is that you can't train loras on it. Most of people use 8 steps but recommended is 12 or 14 steps (yes, there's a huge improvement in quality from 8 to 12).