Home Artists Posts Import Register

Downloads

Content

I decided to kill a huge amount of time for tests. I made a lot of versions of the same lora with different settings in order to understand how to remove artifacts at high resolution and the result drove me into a stupor - there is practically no dependence on the weight of the lore and settings, each model can artifact at a completely different weight, and whatever I do, this is inevitable(( But at least I have identified for myself adequate types of settings for different situations, although I could not solve the problem.

512x768, Hires fix 2, Latent (nearest-exact), Denoising strength 0.5 were used everywhere

I will attach all versions of this model, maybe someone will come in handy. Push

mazav1 - my standard settings, which I used quite often, although not always, there is clearly a slight "roasting" closer to weight 1, but overall quite stable

mazav1A - the most unlike the original set of images at low values, but the most stable at high weight

mazav1B - an experimental model, which, on the whole, partially justified itself, the average noise is the same as that of the fourth model, the accuracy is high and fairly stable

maza1C - just a test that turned out to be good results, high accuracy, good contrast

mazav1D is an unsuccessful model from my point of view, a lot of noise, colors are faded, I don’t know the accuracy, visually here each model in fact needs to be examined with a magnifying glass.

  I would choose 1A or 1C, but just 1 and 1B do not have their advantages and disadvantages, you choose anyway (goodbye 1D, you are the weak link)

P.S. don't ask about the fourth comparison)

Files

Comments

No comments found for this post.