Enter the Five Tag Challenge and have an opportunity to win up to 50,000 buzz from me! A fun challenge I created to see what creative things people can make with a limited amount of tags and a way to continue paying back the community for all the support.
I have created some articles comparing v0.01 to v0.02 so it is easier to see the improvements and changes I have made (one for PG and one for everything else):
https://civarchive.com/articles/29602/pg-comparing-hyphoria-v001-vs-v002
https://civarchive.com/articles/29604/xxx-comparing-hyphoria-v001-vs-v002
The long awaited new version of Hyphoria is here, after almost a year of waiting, I present v1 v0.02!
When I made the first version I was unsure of its quality or if it would even be worth using, but you all really have seemed to enjoy using it and doing things well beyond what I would have imagined with it. I'm glad everyone has loved it but that also made it hard to come up with something to improve, there were always things to improve but with merging and even training it isn't easy to make something that is a categorical improvement at least on my limited budget and available time.
So after experimenting with other models (my real version on Illustrious and my Qwen Image model) I learned a ton and over the last half year I experimented and tried to find something that I liked and felt like an improvement while not being a step down in most ways. I wanted it to feel like the same model but better and that is where v0.02 came to be.
It has much better contrast compared to v0.01, which was washed out especially when you did img2img or hires. The background details often were non-sense or didn't resolve to anything useful, v0.02 in a lot of ways improves on that but can still be nonsense but that is SDXL/Illu for you.
The other major area of improvement was around higher resolution stability being able to reliably generate at 1.5 megapixels with little to no anatomy issues (at least compared to v0.01) and that translates to better stability when doing hires passes on images. In my testing I rarely see anatomy issues like extra body parts even after multiple hires passes (but the denoise and resolution settings you choose will affect that obviously). It should be noted that ultra wide resolutions or high resolution wide aspect ratio generations can have issues with duplication or anatomy warping (e.g. subject laying horizontally) for both versions, portrait aspect ratios are more stable and support higher resolution without issues in v0.02.
v0.02 Recommended Settings
Steps: 20-35
CFG: 3 (2.5-6)
Sampler: Euler a OR DPM++ 2M (Should be stable on most)
Scheduler: Should be stable on most
Resolution: 1mp to 1.5mp
Prompting: Less is more, I often don't even use quality tags outside of
worst qualityin the negatives. Quality tags can lead to a glossier and less flexible result, but sometimes more attractive poses and lighting, experiment.
v0.01 Recommended Settings
Steps: 20-35
CFG: 3 (2.5-6)
Sampler: DPM++ 2M SDE (or DPM++ 2M on CIVITAI Gen)
Scheduler: Karras
Resolution: I recommend 1024x1024 - 1280x1280 for normal generation. Supports up to 1536x1536 without breaking, but can suffer some body extension. Any aspect ratio that equals the same number of pixels should work.
Prompting: This model listens to your prompts, like really listens, so avoid the fluff and keep your prompts focused.
Recommended Positive:
masterpiece, best quality, absurdresRecommended Negative:
worst quality, low qualityYou can often omit pos or neg quality tags depending on what you're trying to make, grungy images for instance will want to pull out some of those quality tags. Some of the other quality tags do work if you want those types of images.
License & Use
SDXL - CreativeML Open RAIL++-M
Illustrious - https://freedevproject.org/faipl-1.0-sd/
NoobAI-XL - https://civarchive.com/models/license/1140829
Copied the above license list from Plant Milk's page.
Description
Experimental merge, no real logic behind the merge outside of picking models I liked and some numbers I thought made sense for weighting them. Future versions will have structured and less subjective evaluations to determine more optimal configurations.
FAQ
Comments (141)
Nice merge recipe, I'll be trying this out!
Leave any feedback or issues you have, looking for ways to improve for future revisions. Thanks!
The annotation terms for the poses are quite accurate, so I'm satisfied with that. As for the eyes, using full-body annotation terms tends to result in a higher chance of distortion. Keep up the good work!
That's true in any model, but definitely something I wish to improve. Thanks for the feedback.
Daddy makes checkpoint 😍😍😍
ilu. i'll try since it's from you 🤤😉
Let me know if you have any issues, looking for things to focus on for future versions.
The main issue I often run into with huge merges is that inpainting performance is extremely poor (bad blending artifacts). Inpainting so far has worked pretty good on this, so good work!
This isn't a typical merge, at least not like what most people do. It's using relatively new techniques that I'm not qualified to explain, but are not too complicated. The benefit though is that it keeps the good parts of models relatively untouched and pulls in the better parts of the merge targets. That's the hope anyways. I've been doing detailers with it quite well, using forced inpainting but haven't done that specifically, will keep that in mind for future versions to test/validate. Thanks for the feedback.
Seems absolutely awesome after some early tests. Unique output.
Thanks again for taking a look. I appreciate all the feedback and am glad to see people are liking it and finding a use for it.
Hyphoria is available on site generation this week. I'd love to hear any feedback on it, especially any issues or things that don't work as you expect. It is an experimental merge and I want to iron out issues for v1.
Can't wait to see what y'all make with it.
amazing checkpoint, 10/10 would use again
Awesome stuff. Thanks for checking it out and love all the styles you were testing.
really outstanding checkpoint
Thanks! Glad to see what you made and hope to see more.if you have any feedback or issues let me know.
damn i just created some images, this model is amazing
Thanks for giving it a try, if you have any feedback let me know. I'm working towards a proper v1.
Unbelievable how damn good it is.
Thanks so much. I was honestly surprised myself. I knew what I wanted to make, but somehow hit gold right from the start.
https://civitai.com/posts/17805386
Here's my result, thank you for an incredible checkpoint.
Thank you for giving it a spin, love the look of those mecha.
Your statement that you built this for prompt adherence isn't idle boasting. Just look at this post. Everything here is just your checkpoint and raw prompting; no Loras or detailers were used. Clean details right down to the pupils. https://civitai.com/posts/17698923
Thank you for the kind words and checking it out. I love seeing what people do with it and especially when they realize how powerful these models are just with the words we put into them.
I purposely chose HARD to quantify words for the one-word challenge, and STILL got amazing images out of it. were they accurate to the definition? Mostly not. Were they COOL AF? YES, All the Yes.
That's the beauty of latent space, it doesn't always make sense, but with enough quality training they can give you something interesting out of very little. Thanks for trying Hyphoria and entering the challenge.
pretty great model. it does add a yellowish hue to every image it creates which is a bit of a shame but at least it's consistent so it's easy to work around. it adheres to prompts very well.
Do you have specific examples of where it is doing this? As well as a model that doesn't? Are you doing local gen?
@ecaj sorry for late response but i think it's pretty apparently even in the example pictures you posted. I'm guessing it's just a result of the base art theme having a very warm undertone. Which is not a bad thing, some people may like that.
@penguiny2357 But that's the thing, without a counter example it's hard to know what you're considering a yellowish tone. Considering that I did no lora merging or anything specific for the base style, it is a pure merge so any bias is an artifact of the other models. That isn't to say I didn't have any control over that, but I cannot see what you are seeing and that is why I wanted examples of where it doesn't happen on other models so I can make sure to look out for it.
Wow this model is my new fav.
Thanks for checking it out!
What hires fix models are everyone using? Mine are ancient and I'm looking for some new ones.
What do you mean? I use this model for any latent upscaling, if you mean for pixel based upscaling, remacri https://civitai.com/models/147759/remacri?modelVersionId=164821, is solid and fast and worth checking out.
@ecaj Thanks I have been using 4x-UniScaleV2_Sharp and found the sweet spot.
@asdfjkasdfasdf Someone just shared this with me, I haven't had a chance to try it, but it looks promising: https://huggingface.co/fal/AuraSR-v2
Smooth and nice Lora, easy to use and clean
Thanks for checking it out!
This is now my go-to model. 😍
Thanks for checking it out, enjoy seeing all the things people make with it.
I wish I'd known this model sooner. I've struggling with WAIs for months getting bad results. this one works with EVERYTHING!!!
I know what you mean, once I found models that listened to me, I learned what I was missing. I sought to make something even more for me, and this is where Hyphoria came from. So glad that people are enjoying it too, and thank you for giving it a try.
so damn good
Thanks for giving it a go and the kind words.
This is awesome, one of my favorites.
Thanks for checking it out, excited to see what you make with it.
Thanks, one of my favorite checkpoint!
Awesome, really appreciate the kind words and you checking it out.
Loving this model as well as the Hyphoria Real version. Very nice indeed!
Thanks for checking them out and sharing the kind words!
Best, just best.
Thanks for the kind words! Glad you like it.
This is an amazing checkpoint, indeed the best, but there is an issue. Whenever a blonde or white haired character is generated, they are forcibly given twintails or pig tails with the hair flower ornament. Is it because it's too deeply trained?
I'm glad to hear you enjoy it, thanks for giving it a try. Now about your issue.
So I'm not entirely sure what you're seeing because in my case, I don't see that. Are you using any loras? What about the prompt, is it using an artist style or some other reference that might affect the styling? Finally quality tags can sometimes affect that as well (or embeddings if you use those).
You can verify yourself, do a test of just "1girl, blond hair, best quality" for pos and "worst quality" for neg. Do the same with white hair. Now layer on quality tags and/or loras and see what is triggering it.
Ultimately the model itself is not trained, it is a pure merge, across 14 total models, so while there are biases baked in, it is very hard to imagine the case with the hair without something else eliciting those looks.
I don't know what kind of voodoo magic you pulled, but this model is hands down one of the best, if not THE BEST, I've ever used. Period. It can do artist styles so good, better than pretty much any other model I've used before. If this is just version 0.1, I can't wait to see what other versions are capable of! It's become my number one go-to model.
So incredibly kind of you to say. I have enjoyed using it and seeing everything people make with it. It will be hard to top, but I want to make sure the next iteration is as good and improves in other areas. Thanks again!
Oh yeah, I forgot to ask. What VAE, if any, should be used with this model?
It comes with the standard SDXL VAE, any that are compatible with that are fine, down to personal preference. I prefer the standard one as it doesn't affect img2img and as such can be used with inpainting and hiresfix. Alternative VAEs can cause issues there but can sometimes be a good way to do base gen if you want features it provides.
Honestly I don't know why this checkpoint is being praised so much... In my tests, it has the same quality as any other model on CivitAI. It might be because of the style you choose but the contrast/color is weak. This checkpoint doesn't accurately render characters trained with LoRA, unlike other checkpoints I've tested even checkpoints that weren’t used as the LoRA base.
On the other hand, this checkpoint does reproduce the style of other artists trained with LoRA very well, including the lineart, brushwork, shading, etc... Still it struggles to produce vivid colors if the artist style is like that.
I appreciate the feedback, do you have any examples on specific LoRAs that had issues? It would help me improve for the next version.
You're right that Illustrious and NoobAI are great base models and anything built on top of them is similarly great. There are many models that I recommend beyond mine, just usually for things that make them unique and interesting to use. At the end of the day, people should use what they enjoy using and works for them.
If there's anything I can do to make Hyphoria work better for you or anyone else I am always willing to help as best I can. Thank you for trying it out.
higly better in draw mode
I'm all Euphoria for Hyphoria. Beautiful work. Awesome results. Thank you a lot!
Thank you for giving it a chance and the kind words! I have loved seeing what everyone does with it.
Ruined me for other models. There is only Hyphoria and Hyphoria Real now 😱
Is good, I think is perfectly balanced in different contexts, the quality is exceptional. Keep going. Thank you.
Works great with styles of already known artists. Especially those that are more stylized. Other models usually do something like 2.5 D, changing the style too much for my tastes. Hopefully if a new version comes out, it will work well with styles as well.
That is one of my goals, I won't release a new version that will get in the way of trained styles/artists.
Best checkpoint ever. Very accurate. Now using only Hyphoria and "real" version.
New to generative AI but very happy with this one and the semi-realistic style.
magical 10/10 A+
fully agree, it cooperates so well with so many loras. The eyes are more stable and retain their sparkle more than many other Illustrious checkpoints out there.
so good, i love it
Would it be possible to train the color/brightness tags from rouwei into this? So we can get true dark scenes in a non vpred checkpoint etc. My only real gripe with every illustrious checkpoint outside of Noobai vpred.
It's interesting, I did merge in rouwei but the way I merge I can't guarantee certain behavior/knowledge unless I validate test and adjust (if that makes sense) and in this case it was not something I explicitly checked for. I will need to look into how it works with rouwei to better understand how it would apply and be merged. I agree having stuff like that is nice so will definitely consider it.
@ecaj Understandable of course if it cannot happen but it would definitely be a nice cherry on top sort of thing as long as it didn't mess with Hyphoria too much.
One of my go-to's if not the default checkpoint I use.
Love seeing what you make, you pull out some of the most creative things from the checkpoints and loras you use. Thanks for showing what can be done.
One of the best I've seen
I've only drawn a few hundred pictures so far, but so far the combination with Lora has been going quite well, and the prompts for not only the characters but also the backgrounds and accessories are accurate.
good job mate
Edit: Hyphoria is back on site! Thank you everyone that helped contribute.
To anyone wonder why you can't use Hyphoria right now, this week has been crazy especially today and I forgot to bid in time. I'm sorry. I meant to put in an automated bid but hadn't. I'm putting in one today so this doesn't happen again.
When it will be back?
@Obrabo Next week, I already put in the bid and have it set to automatic. It is a large bid so there is likely no chance it won't make it.
no worries!
Thank you! no worries
Bro I did the same thing this week too!
@LatteLeopard Damn, well it won't be an issue going forward (well unless things get inflated a ton more) I just feel bad for people that can only use site gen and use my model regularly, feel like I let them down.
Beautiful anime model. Thank you.
Question: Is there a reason why you use DPM++ 2M SDE and not DPM++ 3M SDE?
No particular reason, why do you ask?
@ecaj :D was just wondering, Amazing work on this model btw !
DPM++ 3m SDE in my experience always needs more steps to approach what 2m gives you at far, far lower steps. That being said they're both inferior to DPM++ SDE
Love the skin texture and lighting — especially in complex scenes with 2+ characters. Only issue I noticed: some background elements glitch. Great overall!
Backgrounds are definitely an area I want to improve in the next version. Thanks for the feedback and for trying it out.
What should I use with an artist? For example "by ____" "artist:____" or just the artist name?
Just the artist name.
Why the heck did this get 7 thumbs down, including from the OP?
@clydeleegrahamAI No idea lol
Awsome model it's my go to now !!
are you planing on doing a new version with new charactes westen cartoons ? any time soon ?
I do have plans, am working on the new merge pipeline. I hadn't planned specifically on Western cartoons but I do plan to update the knowledge so I'll try and look into that as well. Timing wise, this is a hobby so I do what I can when I can, so no hard timeframes unfortunately.
@ecaj respect 🫡🫡🫡
Do you know what's up with the bottom and right side of the images? It doesn't happen all the time but it just ignores those edges sometimes. Do you have any plans to fix it either if you ever make a new version?
That is an issue of using a resolution that SDXL latents don't support. I actually faced this issue and prevented it by being selective of what resolutions and aspect ratios I use.
@ecaj which resolutions would you recommend?
@Sfdwackys In general the ones civit has are good staple options, then for upscaling I do two 1.5x hires passes, but you could do one or even a 2x hires. Just have to see which multiple works. I think you need at minimum the resolution to be divisible by 8 on both sides, but sometimes you can still have issues.
Really impressive. How did you knock it out of the park with your first try?!
I did spend the better part of two months prior testing manual merging, then getting clued in on a better merge algo. Learned how to implement it myself after many trials and failed merges (due to the algo being incorrect) and once I got all the pieces in place I was lucky the recipe I came up with worked out as well as it did. So all luck really.
@ecaj I respect the honesty - looking forward to future versions. What are you looking to tune?
: P
good.
Big Thanks for this Amazing Checkpoint :-)
This is one of the only Illustrious models where I can extract a Lora when I finetune it. Almost all other Illus models get stuck at step 65. Any idea what is different? The other models only work if I don't do the text encoder but then they also don't work as well.
I have no clue to be honest. Interesting though.
This is becoming one of my go-to models. Really nice results 👏
0.01? Yeah, right. It's already one of the best checkpoints out there. I find it easy to work with, very responsive to tags - and the output's absolutely beautiful, too.
I've said this before, but when I published it, I was doing it just because many people on discord asked and I was unsure how good it was, to the point that some issues I was seeing (which turned out to be just comfy workflow issues) made me think it was broken in some way.
It turning out to not be broken and being such a solid foundation is partly why I haven't rushed to update it. I want versions to be meaningful upgrades and if I'm being honest it will be hard to top what is already there without some work, which since I made it I've leveled up a ton on how to do things and across many model architectures now.
@ecaj Man, that's so funny and super interesting. I really wonder how many gems and masterclasses in models are dumped because of random unrelated tech issues giving bad results when the truth is it's anything but bad
10/10 💛
This model should be number 1 on the list of Illustrious models
Still an unmatched model for anime style.
"Little bit of this little bit of that" Type model
This model is a stew of many things, and it comes out bloody amazing
This model understands prompts so well and creates some beautiful images. Definitely my favorite so far, does f/f and futanari well(which is pretty much all i go for) Super great!
One of my preferite models! Thank you so much!
Everything I put this up against loses.
One of my preferite! <3
Absolutely great model, probably my favorite IL checkpoint. Thank you!
Any chance of a Klein 9b version since you've made a Qwen version?
I do want to do something with klein but just have not had the time to focus on it. I am definitely very interested though because I feel like robust editing baked in would be fantastic for the local editing experience.
Was great and then all of a sudden it just derailed every prompt i kept trying to give it have never had this issue and all of a sudden your model has a shittier comprehension then SD1
How are you using it? Nothing has changed with the model, so something in how you are using it changed.
@ecaj i found the prompt limitations your model along can make a lot but at a certain point there are more complex positions/character traits it cannot perform and i had to use loras to carry the weight of.
@Crazed_Clara Well those limitations are a limitation if Illustrious as a whole, it doesn't know everything.
I never comment, but this model BLEW me away! It is legit the best illustrious merge out there! GGs!
nice
absolute peak
is the vae baked in?
It includes the standard sdxl vae yes
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.