Thank you, it's a pleasure! And, yes, it's a good time for this question, because the inpaint version will be available tomorrow.

@fp16_guy @axlpr Can someone educate me? I know what inpainting is, but what is an inpaint version of a checkpoint? My googling just brings up inpaint in general, and I can't see a difference in the sample image! Thanks :)

@revilorevilo, it's simple - any inpaint version is based on the model originally made for inpaint. In other words, it makes inpaint better, because it better "understands" how to combine layers, where the end of one object and the beginning of another, and so on. So if you want to make inpaint more efficient, use the inpaint model.

In the preview, inpaint is made with two vertical stripes on the left and right sides (pay attention, for example, to the car on the left).

@fp16_guy I see (I think :) ) - I'm still learning so much I've not looked at inpainting much yet. That's a job for another day I think!

I was going to ask the same question about inpainting. Your explanation and pointers on what to look for cleared things up so much for me. Thank you!

Thank you, if you train something on this model, let me know, I would like to test the result!

The database is too diverse, so there is no single answer. I like the stability of 600x800, but everything that I used for previews is suitable for use - that is, the range 512x512-768x1024 is quite usable (but as always, stability decreases with increasing resolution, there are no miracles here - the base in the form of 1.5 sd affects).

@fp16_guy my preference is 768x768 and it randomly generates mutation and non-realistic pictures, I'm wondering should I change the scale or avoid some poorly trained prompts, like Asian, catfish, cambridge station these sort of things...

@sgmgong28901119, prompt and negative always play a huge role (and if model is flexible, this role increases even more), it is also important to be able to adjust the parameters on a specific image, even a simple cfg tune in the range of 0.5-1 values can significantly correct some shortcomings.

Each model has its own "understanding" of tokens and applies them with different strengths, therefore, adjusting the weights and, in general, using tokens wisely will always improve the result, PicX_real fully complies with the "got what deserved" rule regarding the efforts invested in creating an image.

I can also separately note that many people thoughtlessly throw a huge set of random words and embeddings into negative - this is bad, it kills the diversity of the model, and if prompt is also written lazily for 20-30 tokens, then there will be no good result, this is not an overtrained model working on one style to do everything by itself with it, you need to understand what to write and for what.

Honestly, I tried to analyze sdxl and draw some conclusions that would help the release of my sdxl model, the last time it was about a month ago. But each time I was disappointed - the variety of models is too small, and among them I did not find any that would really interest me at the moment, especially considering how much slower sdxl works.

Perhaps in the future, this will change through the efforts of the community, or we will get an alternative in the form of sd 1.6, but so far there is what there is, and this is not enough so that I can do something pleasant and different from what already exists.

please forget everything about SDXL and get over it it's bullshit. Original SD it's the GOD of all diffusers, you can training it forever and beyond.

@StableDifference Just your opinion. A bad one.

If any model is at least 2 gigabytes in size, it always has VAE. There are 840000 built into this model.

@fp16_guy I'm still getting discolored images with no detail, am I the only one?

@alelimo90985, as you can see from the comments and images, yes. But if you want to figure it out, then you should start with an example of such an image.

@fp16_guy unfortunately I just tried everything, I even removed everything from the invokeai folder except picx_real, copied every setting from a couple of images from the examples but nothing, still flat discolored images, I'm afraid I won't be able to use this model :(

@alelimo90985, this is clearly an incorrect operation of web-ui, I can say for sure that there are no problems with this locally in automatic1111 and in online generators on civitai and tensorart. Have you tried to connect a separate VAE file manually?

there advantage using lcm model , or any regular model with the lcm lora is just the same ? what differences in quality and time ?

@amazingbeauty, I have not conducted such tests at the moment, but I assume that this is a matter of how to use it, and if you are already using LCM-lora, then there is no point in changing it.

I can only remind you about the areas of application of LCM and the reasons why these versions are needed in order not to repeat, I will leave a link to my previous explanation here.

But this is an auxiliary tool for certain purposes, and not a replacement for the main model - you should not equate them with each other or confuse them with each other, because they have different fields of application, and the regular version was and will be the main one for basic generations.

@fp16_guy ty for the patient response :p

There is no "best" or "magic" prompt - these are myths and legends, any image can be obtained by a variety of prompts - they can be different in terms of structure, complexity, and even stylistic descriptions. I do not recommend using a simple prompt, as well as a prompt that does not contain what you want to see.

But each advanced SD user has his own approach, and his approach is developed through trial, error, and building up his own experience, this is a completely unique and individual way (this applies to any model). You can see examples of my prompting, if you want to learn from my experience, on my page all the latest images (except three with animals) are made with PicX_real.

You can also read a little more about how it is worth and not worth making images (in my opinion) here.

@hxra21 This prompt requires specificity, and it plays well with most LoRAs. Make good use of negatives, and push for what you want to see,

@Suppressor @fp16_guy thanks so much for your amazing answer guys!

It's cool that it suited you so well, and if you share a few images in the future, it will be even cooler! Thank you!

I'm glad you like it, but so far if I'm thinking about updating, it's only PicX and PicX_cute, I got a lot of interesting output when creating PicX_real, and I want to use them.

As for PicX_real itself, not much time has passed, and since I literally used the best parts of the best models, maybe now I just won't find any new parts to update the model. I would not like to become an author who makes 10-15 identical versions with different numbers, so I like the idea of updating the model, but only if this update is noticeable and will not be a step backwards.

@fp16_guy that's fair

@fp16_guy I dunno if it was intentional but PicX Real could really just be called PicX, cause it can do like basically anything at all really well, including 2D art and stuff, I find

@diffusionfanatic1173, there is some difference in understanding certain concepts at the level of the model itself (without adding lora and the like). PicX_real interprets an unrealistic concept (for example, hair made of fire) in a more realistic way, but the original PicX will do it differently, since in it parts of realism are adjacent to parts of less realistic models.

PicX_real was assembled only from models that were created for realism, so it is deservedly "real".

Haha, take breaks every hour and walk for at least an hour a day so that no one accuses me of killing people with chicks.

Thanks!

In my case, this is the power of complex merging and extended testing, so training was not applied here.

As for objects such as teeth, I agree that the stability of some parts is not the strongest side of PicX_real, but this is the price for that very flexibility and balance. Perhaps, over time, I will find a way to smooth out the disadvantages a little without losing the advantages, but this is a difficult task, and it cannot be solved quickly.

@fp16_guy Only merging? Now I'm even more curious... I mean usually I train on my own merges and so I know a little bit, but damn, I have never been able to get this level of quality out of it. Good job!
Currently I am working on face pictures. One goal is to get a good variety of types expressions and angles.
You know more than the typical "stare into the cam". And while I am at it I am recovering old pics or just take rare and pixelated screenshots and see what I can recover. Maybe I could send you the result of what I am doing and you can see if you can merge it into your model. If I ever get done with what I am doing... lol
I have also tested to train hands and got some promising results, but that has been a while ago already and my work was interrupted by unreliable versions of dreambooth... A day has only 24h, that's not much for one guy and stuff is changing faster than I can keep up with. :D

@mad_rooky, I don't have much time yet, so I recommend testing this on different scenarios in advance, often during merging something starts to go wrong. If after such testing you will have the results in the form of checkpoints/ lora, which you want to share or discuss something, you can contact me via chat or telegram.

And I can't disagree about 24 hours and the fast-changing world of AI, I have several projects in my folders, but I need to work hard and finish at least one, lol, before they all stop being relevant.

Thank you, cool cat with glowing eyes!

It's also straight up a cleanly-styled 2D anime model if you prompt for it. And a 3D CGI one. And so on... lol

@diffusionfanatic1173 , still, to be fair, I can say that anime models are much more suitable for anime images. Now I don't have time to do this type of model, but fortunately, this niche is never empty without me.

You've earned it!

Nice! can't wait to see what else u can make

I've already answered a similar question, you can read it here.

ayy i got featured

Maybe if a few people show their real interest in this on ko-fi.

@fp16_guy Well you know Im interested, hopefully enough people will come to make it happen :)

I'm interested in an SDXL version of this.

@fp16_guy yes, please! an SDXL model would be great! 

It's all nice, guys, but the main part is that after more than six months of working on models, I want to see not only comments in support to stay motivated, it seems to have gone unnoticed, so you shouldn't hope for SDXL in this way. Similarly, with frequent updates/releases of new models, alas, I cannot endlessly devote a lot of time to work that cannot give me anything in return.

@fp16_guy?? it's not enough to know you could be fully brand your name in such a huge marketplace?

@fp16_guy ko-fi is worth time? this is a new thing for a lot of us,maybe approach it differently. instead instruct us what to do even, if you build it they will (((cum))) for this reason

What benefit does SDXL have? speed and accuracy?

The focus was on that. But now new PicX model is about 60-70% ready (I'm still not sure if it will be a new version of PicX or PicX_real), and it shows very interesting results, already defeating PicX_real in many ways.

@fp16_guy damn wish my pc could handle that. mine generate a photo in 5 minutes

@pornoooo, well, you can try tensor.art, maybe you'll like it, it's still faster.

@fp16_guy Will it be based on SDXL?

@rtpenn34971, nope. I do not use SDXL, so the appearance of such a model will be possible only if there is a great interest in it from other people - for this I created goals on Ko-fi/Boosty. At the moment, I see that there is no interest, so it would be doubly strange to do something that no one really needs.

It's cool that you're so passionate, thank you! I'm preparing an update, it will be different from the current version, but it will definitely be interesting.

@fp16_guy Can't wait to spend hours and hours on it ^^

That site is censored to the max! Waist of time.

@john_wil111, everyone has different usage patterns - so this site is not bad for those who like a simpler interface and who are not focused on nsfw content. And I can also say that they are great guys.

Looking forward to the 2.0 model, tipped everything I have. Im an open source absolutist so I say just release it. Let me know if there is a way to tip you real money through paypal and ill do that, I dont pay for software out of principal. I find myself always coming back to picx when other models fail, excellent work.

@cwpoulsen564, each of my model pages and my profile has all the links necessary for someone to say "thank you" to me if they want to do so.

still waiting and excited for sdxl I just started using it, I personally find this ckpt to be the best one imo, can you make a pony version and the regular sdxl1.0 version?

@Prinsessen, this.

@fp16_guy 😭

I Want to have soft and fair skin, face with Eastern and Western advantages, and prominent nipples

And 400k on my account in total. A big milestone, hopefully 2.0 will help me move on.

How much longer will we wait for 2.0?

@cwpoulsen564, I'm doing this 24/7, that's all I can say now.

Hey bro, why were the cleansed fp16 models taken down? Those were good backups and models. Also, why take down all old Picx models if working on a 2.0 version exclusively for other sites? I tested Rendernet, but it's not as flexible as Civitai. Additionally, Picx had some of the best images on the site. Can you please restore your account content? It was pretty valuable for the community. I will try to support the external new models, but there was no point in taking down the old models.

If you're talking about online generation, it has nothing to do with me, it's a civitai service.

Well, I'm definitely not dead, although I do have some issues I'm dealing with.

Civitai added another one - at the moment I don't know anything about why or what's going on, but I'll let you know when that changes.

The process of removing the ban is being discussed. The reason is my images, and there is nothing to add here, everyone can draw a conclusion for themselves.

If you need some of my content - my HF, for communication - my TG.

Hey man, just saw this. Not sure what happened but hope you sort things out. Your model is definitely one of the all-time faves on this site. Also appreciate your earlier efforts to release pruned version of other models.

Thanks for your PicX_Real. Definitely one of the best checkpoints, including SDXL. Good luck!