Hi, thanks for the feedback, glad you liked it. the duplicate is just my loading error, but you can find my other models on my profile. While I was correcting, I also arranged the graphics and parameters to make them more readable and complete. I'm working on implementations of this and other models, stay tuned to preview them.

thanks for the feedback, I'm already working on it, but I'm not getting a well-functioning model. I don't have enough memory to clean up the inpaint. I'm trying alternative methods, there will be news soon, stay tuned!!

Thanks for your interest, as I said I'm already working on it, but I'm not getting a well-functioning model. I don't have enough memory to clean up the inpaint. I'm trying alternative methods, there will be news soon, stay tuned!!

Thanks again for the feedback!! ✌️

The proposed negatives are just a suggestion to strengthen the generation, they are not necessary, the models also work with blank negatives.

For the example prompts, however, yes, it is just laziness, but also to test the consistency of the model in generating the proposed style with prompts that do not adhere to it.

Thanks anyway for the feedback.

While we're at it, I'd like to ask you a question if you don't mind. I set up this hub to encompass my more specific models. Do you think it's understandable that in addition to the versions you can find different models, or is it better for me to create a separate page for each model?? I still have to fix the description, but not everyone reads it. How would you prefer it to be set up for better accessibility? Thanks again!

@Suzanne A model that does everything already exists, it's sdxl 1.0, but it does everything a little badly. towing a model to do both anime and photorealistic is possible but it won't do either perfectly and will be difficult to fine tune. the same goes for other styles or subjects.

For a Checkpoint that does it all, photorealistic, visit my model 'NewRealityXL'.

My question here, sorry if I didn't make myself clear:, is:

I have several non-generalistic models on which I am experimenting (e.g. 'horror comics style', 'photorealistic fantastic creatures', and so on...), I add them to this page together with the other models (APHRODITE, TOLKIEN, RAPHAEL... ), or do I create a new page for each model?

since in the first case only the last model remains in the foreground, or the one that I choose as the last, and I don't know if it is understood that from above in addition to the versions the other models can be found to generate other things, or if they are not noticed. in the second case, however, having many different models, it could be chaotic and many may not be noticed, making many lose the opportunity to try them.

@dedeffa thanks for the compliments!

Tips for upscaling are:

As a template I recommend using '16xPSNR' or '8x_NMKD-Superscale_150000_G'

as a sample, however, I am happy using 'Heun' with the 'Normal' scheduler or using 'ddim' with the 'ddim_uniform' scheduler

for the tile_padding we should open a separate parenthesis because it is a complex world, but in general to go well with everything I recommend a mask_blur of 64, a tile_padding of 512 and a denoise between 0.1 and 0.3.

based on the starting size (1000x1000, 2000x2000, ...) and the upscale size (2x, 4x, ...) the latter must be raised or lowered to avoid artifacts, repetitions and intersections of visible tiles.

@dedeffa thank you for apreciation.
as regards the cfg it depends a lot on the one used in the txt2img, I usually use 4/5, in some cases I needed to go down to 3.

as steps, however, I almost always use 25, but here too, many times I oscillate between 20 and 35, if I don't like the generation performed, but very rarely.

And yes, the generation time with these upscales is not short.

for tiling I use 'linear' instead, as 'chess' is more difficult to control and creates more artifacts at large upscales.
The other two entries are needed precisely for tiling, since UltimateSDUpscaler does not generate the entire image as a whole, but in blocks, or in stripes, depending on whether you have chosen linear or ches.

in ultimatesdupscaler under the 'mode_type' entry where you choose whether to use 'linear' or 'chess' there are the 'mask_blur' and 'tile_padding' entries.

the second item, 'tile_padding' is used to tell it how many pixels each generation block must be large, the smaller the number the more detailed the upscale will be since the part it takes into consideration is small, but the risk is that it generates unwanted things during upscale, for example if it generates a block in a part of the image without details, such as a shadow area or a flat texture, it does not recognize the part of the image to be detailed and therefore generates a new micro image of the prompt provided.

however, at a higher number, it maintains much more coherence in the upscaled image, with more attention to small details and artifacts in the correct generation, but less detailed in terms of pixels if zoomed.

I personally use a 'tile_padding' of 256, on upscales up to 4K, and 512 for higher upscales.

in general, however, always try to use dimensions by which your image is divisible. (if the side is 1000, don't use 192 but 200. if the side is 768 then you can use 192.

the first item, 'musk blur', instead gives information on how many pixels must take into account on each side to be used to make a transition between one 'tile_padding' and the other so as not to have a clear dividing line between a tile and the other. Here which one to use is more complicated, I'm happy with 8 too, you have to experiment a bit and see how you get on. lately, experimenting on large upscales I've been using 64 a lot.

hi @shapeshifter83, yes I messed up. I wanted to try grouping the sub-models into one hub to make searching easier, but the way the pages on Civitai are managed just made it more confusing.

For this reason I decided to bring everything back to the origin and create different pages for the various models.

The latest version of Afrodite uploaded is 2.0, no version updated recently.

I apologize again for the confusion created, but I am still in an experimentation phase, it is better to understand mistakes after having made them, than to avoid making them but without improving.

Good morning @QuinDexter, I'm sorry that you weren't satisfied with the model, but giving 1⭐ because you don't know how to use it is rude.

Saying that 'it doesn't work' when there are more than 10k satisfied users just makes you less credible.

I'm here to experiment and always offer you something better, for free, but I need truthful feedback to be able to do that.

What didn't satisfy you? Did you generate only that image? What prompt did you use? what scheduler did you use? How many Steps and how much CFG did you use? Have you tried changing seeds?

If you can provide me with additional details maybe I can help you. It's not about votes, it's about the lack of respect for creators who give a shit about people like you, who don't try to use their brains before writing a comment.

I hope I can be useful to you with advice and to bring more effective updated versions. ✌️

Hi @vinnnn, I'm glad you like the model, sorry for the confusion, I was trying to experiment with alternative methods to load my models, but it only created chaos during the various updates, so I decided to go back to the initial steps and propose a page different for each model.

The versions are the same, so yes, v2.0 of AfroditeXL is the same one that was present in LNHZ's Hub.

No, it does not require a Refiner and the VAE is already included.

hi, unfortunately being an SDXL finetune, the minimum weight is this. For smaller checkpoints you have to use SD1.5 which I haven't worked on though. A LoRA for generating photographic nude will be coming soon so you can use it with fewer resources and with different checkpoints.

Stay tuned.

Hi @miuimiguel, thank you very much for the feedback, I'm happy that my work satisfies you, and yes, male genitals are a problem that I've had since the first version. I'm working on it, I'll try to make them acceptable in the next version, but I don't guarantee anything, more likely I'll be able to publish a dedicated LoRA first.