or you can use your favorable sampler/scheduler, but lower dmd2 lora strength (i still get good generation using dpmpp_2m_sde/karas with dmd2 strength set to 0.3-0.5), but it also depend on other loras and embedding you're using too. it you heavily using lora and embedding, lower dmd2 strength (even so it will generate unpredictable result and more hit n miss compared to without dmd2, at least that is so in my case... but 32s vs 90+s generation definitely worth the extra finetune and experiments)

you should try it tho, dmd2 lora converges in much fewer step (6-8 steps compared to 40+, and extra 8-10 steps hires fix if using).

https://civitai.com/models/315602/porn-productivitymulti-concepts

if you're into nsfw, that will be great addition, can handle most poses well

@MachineMinded what is DMD2

@felixxauk110 distillation model lora, basically it allows you to generate images in 4-8 steps (LCM or Euler_A, dpmpp_2s_A with karras scheduler, at cfg 1-1.5, lora strength of 0.6-1.0; or dpmpp_sde with lora strength below 0.3)

link to the lora:
https://huggingface.co/tianweiy/DMD2

menyan1381 Any suggestion on upscaling method? Sampler, settings, etc.? Thanks! I use Forge but can translate whatever you think.

civitaithrowaway28398 whatever works... if you're satisfied with the quality and detail, regular upscaler without denoising should be preferred to keep that detail (realesrgan, clear reality, ultrasharp, esrgan, etc). extra denoising (vae encode -> latent upscale -> ksampling -> vae decode only if your image need extra detail or still blurry; otherwise it will just wasting time for something that might not work as expected. for extra detail, you'd want something from dpm family: dpmpp_2s_A, dpmpp_sde, dpmpp_2m_sde, at low denoise like 0.25-0.4. The general idea is to imbue more detail, not creating a whole different thing. keep the prompt same as the original, same seed and model should provide better result) and after that, if still not big/hi-res enough, you can pass it through upscaler. for screen display tho, 1536-2048 is already detailed enough. just be considerate because bigger=longer.

menyan1381 Cool. Thank you. Yes, I usually only go from 1024x1024 to 1.4x or so via high-res fix. It's just been hard using DMD because my normal Remacri or Latent take some adjusting (the denoising seems completely different with high-res fix when using DMD Lora.

civitaithrowaway28398 when it became hard, split the workflow in 2 part, or pass the model directly to 2nd pass, bypassing the dmd2 lora. (a bit problematic if you're using ipadapter, i've gone that path). also, sometimes its better to finish 1st pass early (0.9-.95%), and let 2nd pass finish the job at low denoising strength. Experiment yourself... i'm struggling with dmd2 now with the recent updates... something feels wrong, but can't pinpoint what it is yet; but i'm quite sure that deterministic sampler now work with dmd2 lora.

@menyan1381 how do you bypass the dmd2 part?

well you couldtake a look at natvis or go to chroma

MachineMinded well natvis is a hidden gem, but if you go for chroma, that will be wild

MachineMinded omg yes, a chroma finetuned on biglust dataset would be nice.