you like my content, support me at:
Patreon in here:
https://linktr.ee/Duc_Haiten
I created my discord server. I'm still quite new to discord server management, free to give opinions.
special note: Using the model is very simple, like the scores, but also not like the scores. The idea behind it is training based on various artistic styles of different artists and some unique AI models, arranging them into separate keywords (perhaps I should have made them numerical codes like score_1234 in Pony base, such as style_123—maybe next time I’ll do that). Each keyword corresponds to a different style, representing an artist or a uniquely special AI style. Users can call up just one keyword to bring out that style, but that’s not the main intention. What I aim for are unique style combinations. You can call up multiple styles at once, for example: A, B, C, or B, C, A, or C, A, B.
Any change in the keywords or the order of the keywords will result in different outcomes. You can even use parentheses like (A) or (A:1.2) to increase or decrease the weight of a keyword. This combination and arrangement will turn the initial 55 styles into hundreds or thousands of different styles, and as long as you don’t reveal your style combinations and keywords, it will be a one-of-a-kind creation unique to you. That was my original intention when creating this project.
Read more here:
https://civarchive.com/articles/8523
Because the image generator of civitai and automatic1111 are different, on civitai you need to set the number of steps about 10-15 steps higher than when using automatic1111.
Update on October 16, 2024:
Pony No Score v7, this will be a groundbreaking version that unlocks enormous potential for pony, marking the largest update to date. V7 integrates 55 styles into one model, each with separate trigger keywords, similar to how Pony previously added scores to the model. The goal is to create a model capable of producing a wide variety of artistic styles, rather than being limited to just a few that are difficult to control. Users can apply individual styles or combine different styles to create their own completely unique artistic vision. Once you hide your prompt and data, no one will know which style keywords you combined to craft such a distinct look—it will be uniquely yours.
I will write a detailed guide on how to use the model along with a list of styles here, so everyone can gain a clearer understanding of what I’m aiming for.
https://civarchive.com/articles/8105
Update on September 8, 2024:
Finally caught up with pony base v6 LOL. This is probably the last update, aside from bug fixes. I feel like I've reached the limit with pony v6, and adding anything more might cause issues elsewhere.
Version 6.0 includes some anime data, better adjustments for NSFW, further tweaks to colors and lighting effects, and additional improvements in natural language.
And the most important thing is earning money. Don’t be fooled by how popular this model seems. In reality, I work 15-16 hours a day, without taking a break, just to earn enough to reinvest in research and model training, with barely enough left to live on. So now I want to see how much I can make with this early access feature on my most popular model.
Update on August 8, 2024:
v5.1+v5.0+v4.0=v5.2
Damn I forgot to turn on early access
Version v5.2 works quite well both with and without scores
Update on August 4, 2024:
Version 5.1, a complete overhaul of the text encoder and adjustments to the U-net. Efforts have been made to minimize minor bugs from versions 5 and 4. Once again, the default importance level of the 3D style has been downgraded. I do not want 3D to be added to images if it is not explicitly mentioned in the prompt. Of course, if reducing the 3D style too much makes many people feel dissatisfied, I will adjust it again. So please let me know your thoughts on this adjusted version.
The parameter settings I have tested:
- Sampling method: Euler a
- Schedule type: SGM Uniform
- Sampling steps: 30-36
- CFG Scale: 8-9
Update on July 23, 2024:
Welcome to pony no score v5. I tried to save it until pony base v6.9 was released, but the waiting period was too long, so I couldn't wait any longer and got to work immediately. I need it to have more soul, more artistic depth, and to convey emotions in each created image. I am fed up with the dryness of AI-generated art, which only looks beautiful but feels empty, unlike something truly nurtured from the creator's soul. I don't know if you understand what I mean, but that is what I am trying to achieve. Perhaps the most noticeable change for you is that the 3D style will be minimized, and the 2D drawing style will undergo slight modifications, with colors feeling more cinematic, and the lighting and shadows will be better and more realistic.
Some other updates: improved text encoder, more shooting angles, more art and 2D data to replace Nijijourney data
Recommended settings:
Sampling method: DPM++ SDE
Schedule type: SGM Uniform (this one is a must)
Sampling steps: 30-36
CFG Scale: 9
update May 27, 2024:
This will be the final version until pony base v7 comes out. In addition, the alpha 1 version of the pony-real model has been released, everyone can check it out, but don't expect much from the alpha.
Thank you very much and apologies to Deathless, who helped me find and process around 8,000-10,000 image data for the pony-no-score version 4.0, whom I forgot to mention. So much happened in a short time that it gave me a headache. Fortunately, he reminded me, at least he didn't curse me out harshly LOL.
(By the way, currently I'm doing everything alone again, if anyone is interested and willing to help prepare data for training my model, please feel free to contact me.)
update May 22, 2024:
Final beta version of v4, the next update will be its official version. When the official version of pony no score v4 is released, it will also be the day pony-realistic alpha will be uploaded
updated May 14, 2024:
Version 3.5 for Pony-no_score.
- Edit the text encoder
- Update more 3D image data
- added more realistic skin models for 3d
- reduce unnecessary details that are not in the prompt (creativity is a bit reduced but accuracy will be higher)
- Enhance solid contours
This is a separate version from v4 in training, so I need the most objective reviews.
Update May 7, 2024:
I must say that I am really surprised and happy to see everyone's warm reception for the DucHaiten-Pony-XL model (no-score). Each new version gets more attention than the previous version, it seems I'm still on the right track.
To express my sincere thanks to everyone, I will post the v4 beta version first, everyone can try it out and give advice and suggestions for it. Note that because it is only a beta version, there will be many changes in the future.
Update May 3, 2024:
After hearing comments about v2 seeming to lose the diversity and balance of styles compared to v1, this time I tried to adjust that problem. In addition, a lot of new data has been added in v3. They are arranged to intersperse with Pony's previous data, with the purpose of supplementing and creating connections with each other. I do so with the desire to create a more complete ecosystem for ponies. But because I'm the only one doing it all on my own with such a small budget, I can only try to limit my abilities as much as possible. PS. Colors are more vivid and brighter.
updated April 28, 2024:
The upgraded version 2.0 will have many major changes, first is adding more handsome men data, next is improving weak natural language capabilities, and finally adding more artistic feeling to the model. There are also some smaller adjustments like details and lighting
A model created from the original pony model and combined with my various models. I tried to keep almost all of pony's text encoder and only change the image data, so I recommend you learn more about pony if you don't know how to use "tag score" properly. When using this model, you will feel that the image is less random and don't has too many unnecessary details like the pony, and will feel smoother and cleaner, Eliminates the possibility that some prompts will produce harsh images or burnt colors, the face is made clearer, the hands, feet, and pose are adjusted for fewer errors. But if you gain one, you lose the other. It can only be said that each person has its own style, depending on your preferences.
If you want to use my models to create unlimited images at a small cost without having to install anything, mage.space will be a good choice, i have some models that are exclusively available on there.
Description
FAQ
Comments (27)
I made some of the images you posted for 3.5, using most of the prompts that you wrote. Here's the link:
https://civitai.com/posts/2736416
The first thing I did was, take out all of the Score_# strings in the prompts. Some prompts needed a little help, like the little blue guy had "weapon in his hand", it kept wanting to add a third human hand to the image to hold the weapon. The Jill Valnetine one, I had to make sure to add "1girl, solo", otherwise it kept adding more lady cops. I didn't have to make any changes to the bikini prompt.
I did notice that all the negative prompts have the same first elements. "(worst quality:1.2), (low quality:1.2), (normal quality:1.2)," Which is basically like having "score_6, score_5, score_4" in the negative prompt.
The other problems are kind of the same with most Pony models. Eyes are hard to get right. I really have to dial in my steps and cfg settings. Prompts have to be really really specific, and you can't just write natural english to get what you want. I'm thinking that most any Pony based model won't ever get close to true Natural Language Processing like we all hope it will.
I'd like to see a much wider tolerance for steps, so I'm not choosing between 22 or 23 steps to get the details as close as I can get them to what I wanted. That's what I'd really like to see out of the next iteration of this model.
I'm on a 16-inch, Nov 2023 MacBook Pro with the Apple M3 Max chip, and 128GB of RAM. I installed SDForgeUI from a git pull. That's about it.
Thank you for your detailed feedback on v3.5, particularly regarding the issue of score tags. You will still encounter some problems if you remove them because the pony base intentionally adds poor-quality image data so that adding score tag keywords can improve the images. I don't really want to completely remove the score tag from the pony base because, to me, it has advantages, not just drawbacks. I simply reduced the excessive influence of the score tag so that those unfamiliar with using it can still use my model easily without needing to add score tags if they don't want to.
Secondly, the segment "(worst quality:1.2), (low quality:1.2), (normal quality:1.2)" is not mandatory; it's just my usage habit. Regarding the natural language processing issue and the various minor errors of the pony base, I've tried to fix them, but it's very difficult to eliminate them entirely, and I can only make it better. Perhaps we just have to wait for the release of pony base model v7. It's similar to game mods; they are just mods, and you still need to wait for the game developers to release patches or create a new, less buggy game.
妈的~我的硬盘啊！
Is it not working now? I keep getting a error message and won't generate
are you talking civitai? i don't know if their generator is not working then you can ask admin civitai
@DucHaiten yeah sorry on the civitai generator. other check point works fine only this one say: "Failed to generate, Some of your resources are not available for generation" idk probably it's my a problem in my end, it's a shame cause I love this generator
@GothicAddiction It doesn't work for me either. Unable to generate in versions that were posted before version 3.5. It feels like you can generate in the version that is first in the list or added most recently.
I'm using Automatic1111 and it only produces blue blocks. weird.
@Dionisius93 Well, if this is your problem, it's because civitai only allows 1 version to run, so i always choose the best version, or i think is best version. it would be nice if civitai allowed running all versions in post
@Fal191 Pay attention to the vae section to see if there are any problems, try changing your typing prompt, try running other sdxl or pony models to see if the same error occurs.
@DucHaiten I see now, thanks for the clarification.
same. Totally unable to use right now.
@RobertKing Can only run v3.5 model, pay attention to see if the model has been changed
@DucHaiten On civitai it say v3.5
@DucHaiten You are right the generator was using a 3.0 model. Not sure how to get it to show up on Civitai under the 'Swap' option but when I go to the DucHaiten page and hit 'Create' it uses the v3.5 and works.
This is only producing black squares (with a little bit of color in upper right hand corner) other checkpoints are working.
Other checkpoints are sd1.5 and sdxl or even pony
Well... after some time playing around with v4 Beta, I may say that for me, this one replaces AIart-SDXL for characters and dynamic stuff.
But also, I noticed that it is bad with fangs, hands, gloving eyes, cat pupils, face scales (aka Raen from FF XVI), piercing, modern goth girls (they are all from the 18th century with crying makeup xD), and overall a bit too vivid colors with way contrasting cold highlights - bad if you want to use high CFG for extra stylization as it boosts colors even more.
But I use a bit of Astraali MOGU (around 0.2-0.3) into the mix to fix colors and boost aesthetics just a little.
Hope that this may be helpful... somehow.
Is this good for training LORA? I've been training a men lora with autismpony but this has produced such great results.
Yes
Okay, 3.5 is giving me problems. ;)
I normally setup my prompts generate 10 different girls and pick the best 4-5 face /hair / outfit combinations and prompt them onto the next stage of the scenario, and the next, hires fixing as i go till i have 4-5 different stages for each girl. takes 1-2 hours, and i have a nice round 20 images to post.
But with 3.5? all bloody 10 girls look too good, so i can't pick! So i move all 10 to the next stage/hires fixing, till 3-4 hours later i'm left with an even bigger problem of picking the best 20 out of 50 images, which is nigh impossible because they all came out so well but with their own individual nuance.
Thank you ever so much for such a wonderful problem.
But is it still a problem? okay i will fix it :)))))))))
@DucHaiten it's honestly so good, i've been able to delete so many older checkpoints and lora's because pretty much everything i need is all in your checkpoint, so much HDD space recovered.
(just to confirm, there is no "problem", i was just joking, the checkpoint is perfect, sorry if i wasn't clear, i'm english so my sarcasm Lora is baked in)
@rhult gpt chat can understand your jokes so don't worry
Great work with the 3.5. I didn't know about your work a few days ago. Now I cannot ignore this checkpoint. It's so good, definitely one of my favorites.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.