Trained on Animagine XL V3: https://civarchive.com/models/260267/animagine-xl-v3
Description
FAQ
Comments (15)
Would you consider adding lower rank versions of your style LoRAs in the future?
@hhks Thanks for your support and suggestions!
Training with a lower DIM (for example 32) takes slightly less amount of time but gets worse in learning artist style. If you need to reduce LoRA size, try "remake dimension" using extension SuperMerge.
Here is comparison of:
1. anmiXL_V1 (870MB)
2. anmiXL_V1 remake dimension to DIM32 (217MB)
3. Same dataset trained with DIM32 natively (217MB)
https://civitai.com/posts/1379092
I personally feel "remake dimension" is better than training with DIM32 natively.
Your 1.5 artist lora has always been one of my favourites, yes, 1.5 lora.
But your XL artist lora I have to hesitate to download every time, why? Have I lost faith in you? Or am I just desperate for XL?
Ah, suddenly I look at my hard drive and I understand!
Because your XL artist lora's capacity volume is all of 870.32 MB, which is very unfriendly to hard drives with no space. Can you make it smaller?
--Translation software
@aertt6950 Thanks for your support and suggestions!
Training with a lower DIM (for example 32) takes slightly less amount of time but gets worse in learning artist style. If you need to reduce LoRA size, try "remake dimension" using extension SuperMerge.
Here is comparison of:
1. anmiXL_V1 (870MB)
2. anmiXL_V1 remake dimension to DIM32 (217MB)
3. Same dataset trained with DIM32 natively (217MB)
https://civitai.com/posts/1379092
I personally feel "remake dimension" is better than training with DIM32 natively.
@silvermoong The difference is pretty big when you look at it closely, got it.
BTW, can you make freng's 1.5 lora? You've only made the XL version of this artist. Of course here you can find another freng 1.5 lora, but hopefully you can make one too. --Translation software
@aertt6950 Added freng 1.5 into my to-do list. But my SD1.5 training methods takes way more time and labor than SDXL. I probably won't do it soon.
@silvermoong Thank you. And Happy Chinese New Year. --Translation software
could you retrain this Lora with DPO? XL models are being a revolution with DPO technology and can be used for both Models and Loras,the styles trained with DPO are more consistent and solid when using prompts,I will leave the information I have compiled with a copy and paste from previous post for you to read and update, I am excited to see what Loras and model creators can come up with in the next few months,as I see you used Animagine XL V3,it is a good model but instead I recommend Comradeship XL,it has DPO,try it with prompts and you will see the magic,you can replicate hundreds or thousands of characters perfectly without the need for loras.
(copy and paste)
here is a list of all the models that I have seen and tested that have DPO
1-AutismMix SDXL
https://civitai.com/models/288584?modelVersionId=324619 (all three of their model mixes have DPO)
2-Confetti Comrade Mix
https://civitai.com/models/280344/confetti-comrade-mix(here I had left some comments giving ideas to move the community forward, anyway I will leave my ideas in this post in case someone reads it and wants to know more about it)
3-Js2Prony
https://civitai.com/models/280321/js2prony
4-Comradeship XL
https://civitai.com/models/246299/comradeship-xl(my favorite so far)
5-[0.8(ponyDiffusionV6XL_v6) + 0.2(reproginexl_v10)] > NAI 3 !
https://civitai.com/models/263797/08ponydiffusionv6xlv6-02reproginexlv10-greater-nai-3
6-Hassaku XL
https://civitai.com/models/140272/hassaku-xl-sfwandnsfw-beta(this model understands prompts well but is not good at replicating characters without loras,the other models in the list including the Pony Diffusion V6 XL can replicate characters perfectly,even better than Loras already made in the civitai community)
the list I just left have good implementation of (DPO) the idea of this post is that the creators already famous in civitai begin to implement this technology,98% of civitai models are mediocre but with this technology the existing models can make magic,genitals,legs,hands,expressions,poses, concepts etc make more sense with these models already named(Pony Diffusion V6 XL was the first to implement the technology to a XL model) but the ones already named here are of very good quality also it would be great to continue seeing models with many different styles and more polished in the future) researching I discovered that if you mix a model that already has(DPO) with another model already existing civitai gives results like these. ..the technology works very well with XL but as far as I can see very bad with 1.5(for the moment) I think the future this year is XL I think the other model and lora creators should all work for XL,it feels good when you put indications and you get almost everything as you want,that with 1.5 is impossible,I am using XL,without controlnet and without Lora I have had incredible results,the mixes with the DPO models I think it is better to do them in Comfyui,I hope this post can help you,there are many people who still don't know about the technology
I took the time to test about 50 XL models and all were exactly the same as the 1.5,bad anatomy etc,in search of DPO models I did my tests by replicating famous characters,an effective way to test is using as a prompt:
cammy,chun li,jinx,league of legends,etc,if characters like that replicate perfectly the model has DPO,if the replica of the character doesn't come out then it doesn't,the models in this list can replicate the characters already named and many many more,(just do your list of character tests)....
PS:there are models in civitai that claim to have DPO and don't have them or are very bad,all of them were tested by me so I save you time
here you can read the post of this page to know a little more about DPO
https://civitai.com/models/242825/dpo-direct-preference-optimization-lora-for-xl-and-15-openrail
@FireBlack85 Thanks for your suggests!
I read a few DPO articles. DPO seems to be a way of refining checkpoints and mhdang released examples fine-tuned SDXL 1.0 and SD 1.5. So a DPO checkpoint could mean for one that is refined with DPO method.
On the other hand, I found two LoRA versions of DPO. The first says it is extracted by (DPO SDXL - SDXL)
https://civitai.com/models/242825?modelVersionId=273996
https://civitai.com/models/237775/sdxl-dpo-turbo-lora
Comradeship XL's description shows it merges a DPO LoRA. So a DPO checkpoint could also mean for a one that merges the DPO LoRA.
It could be beneficial to train on a refined DPO checkpoint.
However, I'm not very confident for the LoRA version. I will test the two DPO LoRAs, merge the better one to Animagine XL V3. I will re-train a few style LoRAs and compare for improvements.
After that I will try training on Comradeship XL.
@silvermoong most of the models are being trained with Pony Diffusion as a reference,there are few models at the moment that have good DPO implementation,but you can do your tests,you can test the DPO models and Loras separately so you can test the prompts and their efficiency, I tried to test your Lora with the DPO model I mentioned and it doesn't work,in fact it didn't work with any of the DPO models I have,I like the style of hews and others out there that don't have DPO yet,if I could compare DPO with something else it would be like rollback in figting games,I hope your tests go well
@FireBlack85 I started a test with Yoneyama Mai dataset.
The capability issue might not be DPO specific. LoRA seems to be powerful enough to work with any SD1.5 checkpoint. However, SDXL LoRA may only work well on its base model or one that has merged this base model.
@FireBlack85 I just finished a quick test. Training on "Animagine XL V3 + DPO LoRA" gives me worse results. Which is actually expected. I never had good luck training on LoRA-merged checkpoint on SD1.5.
Which leads to the question: is there really a checkpoint refined with DPO method? I mean, if someone refined his checkpoint with DPO, he would definitely mention it.
@silvermoong here you have an image of what this creator did to make his model work with DPO https://civitai.com/images/6319999
looking through the few DPO models you may find something that can help you.
here he mixed 0.8 and 0.2 of a model that did not have DPO as a result this model came out.
the community is creating styles with DPO little by little.
if you need any help I leave you this profile
https://civitai.com/user/bakariso
all those Loras have DPO, maybe if you ask him he can help you.
@FireBlack85 I see. Seems like ponyDiffusionV6XL is the one trained with DPO. It came with the anime style comparing to SDXL-DPO.
I got DPO worked on Animagine XL V3 by using a DPO LoRA. It improved the outcome.
For [0.8(ponyDiffusionV6XL_v6) + 0.2(reproginexl_v10)]:
ponyDiffusionV6XL_v6: unknown, the owner didn't mention anything about DPO.
ReprogineXL: is a merge of Reproduction_LCMv2, animagine V3 and DPO LoRA
Reproduction_LCMv2: is a merge of Reproduction and LCM LoRA
Reproduction is a merge of animagine V2 and something else.
The above merge uses:
ponyDiffusion
animagine V2 and V3
DPO LoRA
LCM LoRA
I will test these out seperately.
Tested:
LCM is trading quality for speed.
Merging DPO checkpoint into animagine V3 using the same method in Reproduction_LCMv2 produces worse effect than using LoRA directly.
Training on ponyDiffusion now
Thank you for the wonderful Artist XL Lora.
Can you do me a favor and create a style Lora for azullane's illustrator, soaryuna （侑了个侑） (https://weibo.com/2277541084/JB0Kwy3Py). I think he is the best painter, although he is not well known to the public.