If you like my content, support me at:
https://ko-fi.com/duchaiten
Patreon is here (but there's nothing in it):
https://linktr.ee/Duc_Haiten
I created my Discord server. I'm still quite new to Discord server management, feel free to share your opinions.
https://discord.gg/vKEW6jqa49
PLS Read this article to learn how to use the Snapvu model.
After half a year, I finally completed my base model, DHXL. I can say that I was the only one who put a huge amount of time, money, and effort into it. DucHaitenXL is trained from IL 0.1, meaning it's based on the SDXL architecture but with a higher image resolution (1250x1250 at 1:1). DHXL was trained on a massive dataset. From my previous experience training PONY models, I separated different styles into 49 distinct keywords, from style_0 to style_48. However, unlike PONY, these are not quality-based keywords like scores; they are just priority markers for a particular art style. You don't have to use them if you don't find them necessary. I might later write a detailed guide explaining the specific effects of each keyword, but that will be in a separate post. With 49 style activation keywords, they can be used individually or combined to create unique new styles. However, it’s best to combine only up to six styles—any more, and stability decreases.
Detailed instructions on how to use the model:
https://civarchive.com/articles/11924/introduction-and-user-guide-for-new-model-base-dhxl
https://civarchive.com/models/1291404/dhxlv10
list of styles:
https://civarchive.com/articles/12196
Example:
"style_12, style_9, style_22, style_8, style_24, style_37, BREAK"
Some might ask, "How is this different from LoRA?" The difference is that LoRA cannot be used with too many at once. It forces the checkpoint to follow its structure, while this method works as a mutual compromise, allowing for a shared understanding between the base model and the user-defined styles.
Recommended Settings:
DPM++ 2M
SGM Uniform
50 steps
CFG 8.5
2D-2.5D Negative Prompts:
(worst quality:1.5), (low quality:1.5), bad anatomy, bad or heavily deformed hands and fingers, extra limbs, missing limbs, conjoined fingers, signature, username, artist name
3D-Realistic Negative Prompts:
(worst quality:1.5), (low quality:1.5), (normal quality:1.5), lowres, bad anatomy, bad hands, multiple eyebrows, cropped, extra limb, missing limbs, deformed hands, long neck, long body, signature, username, artist name, conjoined fingers, deformed fingers, ugly eyes, imperfect eyes, skewed eyes, unnatural face, unnatural body, error, painting by bad-artist
Description
FAQ
Comments (43)
DucHaiten is one of the most beautiful pony models, looking forward to learning the IL version! It seems to retain artist tags from the IL base, so much potential.
Other competing models support vpred, but what are the plans for support?
I completly understand the idea behind tagging this model as being SDXL1.0, but this prevent using Illustrious LoRA in the Civitai Generator (not that i care too much, i generate picture locally)
I’ve only briefly tested it and will test it more tomorrow, but the ability to pick styles is really cool. It’s like having multiple finetunes in one model. Instead of needing to download several models, you can have one model that does everything. From my very brief testing so far, it seems really good.
(style_1, style_9, style_38, masterpiece, best quality, amazing quality, very aesthetic, absurd res, scenery).
Works well as a positive prompt.
CFG can have a large effect on how the styles turn out. So I'd suggest experimentation.
Can't wait to try it ! Any VAE you recommend ?
don't worry about vae, it's already in the model, you just don't need to add another vae
@DucHaiten I tried the model and it's incredible. Would you recommend training lora directly on it for adding another style or character ?
@Ggrue i will definitely train lora for it, but i need a little break after putting so much effort into this model. also i want to focus on making this model more popular so civitai can create its own category
@DucHaiten Of course. I was asking that if I want to create loras for it, should I use it for base model ? Any tips for adding a style ?
@Ggrue like normal lora sdxl train, just change base model to DHXL
@DucHaiten Thank you, I will try
@DucHaiten Any way we can help make it more popular?
@kunde2 yes pls
DHXL 1.0 - Style Overview Here
Style_3 is on par with realistic Pony and Illustrious finetunes. Every lora works.
Big fan of you realistic Pony version, this one looks great as well, just wished was based on Noob rather than Illustrious 0.1
I have not used Noob, can you give some comparative reviews between Noob and IL, what are the differences and advantages? Please be objective.
@DucHaiten 1. Noob is trained on the entire Danbooru AND E621 dataset - w/ some photoreal dataset.
2. Noob is certified V-pred model meaning a larger color spectrum with deeper blacks and lighter highlights, and the other benefit of V-pred which is prompt adherence.
3. Trained to understand natural language prompts as well.
@DucHaiten
https://cdn.discordapp.com/attachments/1179853857174593646/1339588511640322178/ee1506ce0a1a04c06cea060fbf6410e1a3d8d9719c17130cc5f02336e635f8a3.png?ex=67c1b985&is=67c06805&hm=1c10956e941f3aa39e31e3794f32641aea88d89fa22a1d9914cdaf31abe20eac&
Illustrious 0.1 was trained with 7.5 million pictures, Noob eps added 12 million (total 19.5, almost as Illustrious 2.0) and Noob v-pred added 16 more million on top of it (reaching 35.5M).
This just about the size of the dataset. but Noob (for 2d) give me personally better results (IL looked like old stuff to me, NOTE: your previews are fantastic, nothing to complain about).
It has great knowledge of characters and actions, you can have different characters without having them mixed (talking about anime/game characters) I DON'T know if Illustrious is able to do the same to this extent.
From what I heard finetuning Noob v-pred is rather difficult so is probably better using eps, also there should be a newer version of Illustrious (released not too loong ago)
@DucHaiten Noobai also has a deep knowledge of danbooru tags, sometimes even understanding concepts and characters with image count as low as 50 images. It knows nearly the entirety of danbooru's artists and characters up to december/2024 (if memory serves me).
Once you get the hang of a Vpred model, it's hard to go back to Eps and it's hallucinations.
@Churrus Too bad I found out about it too late, just because pony and IL are more famous
@DucHaiten Seems to be a common occurrence... Your model is great, cheers!
hi ! thanks for your work. i got some questions :
- we have to generate with 1250x1250 resolution ?
- clip_skip = 2 ?
- for training on photos, tagging style_2 is correct ?
- for training a character, "triggerword, 1girl" is ok with a keep_token = 2 ?
bye :)
1. 1250x1250 is the largest resolution the model can produce with almost zero error rate, giving a better image, it does not mean you are forced to create 1250x1250 images.
2. correct use of clip skip 2
3. i don't understand what you mean
4. That depends on whether you want to add keywords in addition to triggerword to determine the default element for lora or not. For example triggerword, 1girl with keep_token is 2, lora will default that character to female, even if you set it to 1boy, it is difficult to change.
style_3 is better for realism. I dont know, if you have to tag your datas.
@DucHaiten thanks for your answer ! in, 3. I mean when I'm training pictures with your model, if it is photos (style), is it better to add a tag "style_2" with the other tags, to say the model "hey, it's photo, you know, the style_2 so you know what to do maybe better than without this information" !
@cheskykaras616 Ok, thanks, I gonna give it a try, I'm searching for non 3D photorealism (style_2 tends to produce 3d sometimes).
This is the most coherent realism illustrious model I've tried, so good work.
However, I don't really understand the reasoning behind the style_n tags. Ponyv7 is doing it to create new styles and prevent replicating real artist styles, but for this model it's pretty clear what the styles are based on. I think tagging them as what they supposed to represent would have made more sense.
Agreed, also @DucHaiten will there be any way for us to know what artist which style_x tag is based on?
I love the distinct style of creativity you put into your models. As this would really have to be my new favorite from your entire lineup. At least for me the the "clean/simple" keywords used to style out images coherently from the same model without worry about losing the current prompt comprehension. Sadly it seems finnicky to preserve this in a model merge with my personal ones. But overall I'm loving this and it's quite refreshing. Cheers to you and thank you for your hard work!
Thanks for your appreciation, please read more of my tutorials on the DHXL model to get a clearer view of it from my perspective
It's amazing, choosing a style has never felt so right. As a request, can you add dark fantasy style to the next version?
I have to know what the dark fantasy style you're talking about is, if it's just a form of classic anime from the 70s, 80s, 90s, with a dark medieval style and magic, then I have several different styles of classic anime in there, some are extremely dark, you just need to add some other keywords in the prompt to direct it to the result you want, you can also mix the styles to make it a unique style without looking like the original.
@DucHaiten thx for the message,i just trained Illustrious lora after find this model.I made a selection of dark fantasy themed pictures that I had on my computer that I liked and I didn't think too much about which style.I wish I had read the comment before that.
Thank you for staying at Civitai. Happy to see you back.
can you please add this model to your TensorArt account?
tensor doesn't pay, you can use shakker or mage
Excited to log on and find this. Some of the best creations I've been able to get are out of your later Pony models, and I wondered why we hadn't seen an IL version. So a checkpoint-trained, Illustrious-based DucHaiten model with well-defined style keywords? Wonderful surprise and absolute auto-download.
As a Fooocus user who always goes with 30 steps because it usually works, I found that I actually HAVE to follow your steps recommendation, with 30 steps the result came out raw. Switching to Quality performance for this model.
Style Overview here:
Overview PG rating
Overview R rating
Overview XXX rating
im using this model to train loras in fp8 with my rtx2060 mobile 6gb vram and im getting amazing results with this model i hope you get native suport on civit ai to train loras and generating images here, and when i put the loras on civit feel free to merge to create new checkpoints amazing job
I hope so too, but it seems very difficult with this low download count, Civitai needed this model to become really popular before they would accept it.
Excellent model. Have you considered train a v-pred version?