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
Focused on improving contrast, darker darks and lighter lights, as well as higher resolution stability allowing for 1.5 megapixel base gens without anatomy issues in most cases and sometimes up to 2 megapixels (this also means hires should be more stable as well).
Tried to improve overall details and creativity of the model while keeping its prompt adherence and knowledge, unfortunately I did not have the time to train in new characters or artists and merging models that had that knowledge was not as reliable as I had hoped.
FAQ
Comments (54)
Cult of Gremmy approved
Finally V0.02...
I didn't think I was alive by the time they released v0.02
being gay lord now daddy
Feels like a major step down in style. V0.01 looks much better.
I appreciate your feedback, but I did a ton of side by side testing and there were very few times v0.01 beat out v0.02 for me. I'll be publishing an article detailing the differences and showcasing the effort put into it, but for the time being the old version will be on site and of course downloadable so people can use whatever they prefer.
Fortunately 0.01 is still available for you, then. It's all based on subjective preference and taste, so it's sometimes hard to determine if something is 'better' or 'worse'.
Congrats on the new release, 0.01 was a huge inspiration for a lot of people, so I'm quite happy to see you're still active and pushing for newer releases!
@ecaj Thanks for the response - look forward to the article.
Originally I thought the same, but after modifying lora strengths, tags, and CFG/Steps I ended up clearly seeing how much better 0.02 was. This version seems to have very strong adherence to what you put into it, so generations become overly saturated, soupy, etc. Lowering what you usually would do in 0.01 provided 10/10 better generations in comparison.
Imo, 0.02 is a cut above other illustrious merges. Very good job @ecaj!
it's here and it's fantastic
Thank you for this. Honestly, this is my go to checkpoint, since 0.01. Kinda surprised myself after using quite a lot of other checkpoints to try them out, I always default to this one. It's my favorite.
Going to play around with the v0.02.
Motivated me at 4 a.m. to start exploring. I'm happy about the new version. Thank you!
Nice model but it suffers a lot from concept bleed. I tried adding the "star brooch" tag to a prompt, and it made literally everything in the image star-themed. lol
EDIT: this problem can be minimized through prompting (see replies)
That is an interesting finding, though testing a handful of models using the same prompt as the one on your profile (but with star brooch added) some did not react to the tag at all and didn't add one, or had concept bleed. When I added the tag into a more complex prompt there was no concept bleed. I believe this is a classic case of how SDXL works with tags and while it is amplified with my model it's a common issue.
Regardless thanks for the feedback and trying the model.
@ecaj Yeah, the prompt was pretty short (only about 12 tags), so I guess "star" was having an oversized impact on the image. I tried adding a few more tags to the positive and adding some starry tags to the negative, and that really helped. (I added the result to my other post if you want to see.) Also thanks for the model, the reply, and the tip as well!
Holy Smokes our boy is back!
Hands down the best local model I have used.
This is a really great checkpoint! It gives you excellent results and stays stable even with a lot of Loras.
I highly recommend it! :)
Anyone else having trouble using v0.02 in Forge? 🤔
Edit: Actually works fine in Forge; the issue was apparently due to a corrupted download.
After a few tests, I can also confirm that quality tags aren't mandatory. This checkpoint is simply amazing! ✨
Thank you @ecaj ! 🙏🏻
I wouldn't see why there would be an issue but I don't use forge.
@ecaj I'm not sure either. I've heard that could happen, but I don't know how it works. 😅
v0.01 works perfectly well, but I only get noise with v0.02.
Still, thank you for your reply. 🌹
@K0n3zum1 Hmm, try redownloading it just to ensure it didn't get corrupted. I'll see if I can find anything on that.
@ecaj Okay, took me more time than I would have liked (between Civit's sudden deaths and a nasty wrist tendonitis), but I was finally able to redownload it, and it seems you were right; it probably got corrupted the first time, but now it works just fine! ✨
Thanks a lot for your time and help, it's super kind of you! 🤗💕
And thank you for this amazing checkpoint! 💖I can't wait to try it further! 🥰
Come on, forge is outdated, switch to Invoke
I can say that I love your latest version of Hyphoria 💕💕💕
You: you don't need any quality tags except 'worst quality'
Me: no way that's true
me tries it
Me: holy shit, this works really good!
v1 was already my defeault go-to next to Vixons' models. v2 is just that, but better. Your and Vixons' models have some magical ability to 'stabilize' Lora's, next to great prompt adherence, which makes the model not trip over its feet when using certain Lora's or many of them.
Insane model.
Should I use Hi.res fix with this model? If yes, what settings? (upscaler, hi res stepes, denoising)
In general it is worth hiresfix for Illu/sdxl unless the subject is very close to the camera (like shoulders + head), otherwise details can be under defined.
What settings really depends on what you want to do, there's no best settings as far as the hires process.
Surprised by checkpoint quality, good job!
Amazing model! such a hidden gem!
the best anime model
What's the cutoff date for this model? Or still stuck in 2024?
I did not actively train in more recent data, I also did not have much success in merging newer trained models without degrading quality. I want to train newer knowledge in but still working on the datasets on that front as I would need to source outside of danbooru.
Is there an artist list? I'm trying to get a classical impressionism look.
LONG LIVE THE HYPHORIA!
Welcome back!
I love this model and your five tag challenge helped me to stop spamming quality tags and negatives, and to write more thoughtful prompts
Did you delete the model on Yodalio? I was checking my work and noticed that some of my entries say the model is no longer available.
I have never heard of that platform, so I never put it there.
Great Model always a fan of your work.
does this version know about BREAK or does it have a way to separate charater traits?
BREAK is a feature of software running the models, like A1111 and Comfy with custom nodes. So my model doesn't know anything about it. Separating traits can be done in some ways with different software but illustrious and sdxl as a whole aren't very good at it just from a prompt.
Any Chance for Anima Based Finetune?
I'm considering it, but Anima does so much already that I haven't really landed on what I'd contribute. I don't want to just train on illu output and perpetuate a weaker models issues.
@ecaj nah, just do it if you want, no need to worry about contributing
@Protonium I have ideas for some improvements I could make, but training takes time and I try not to rush out bad stuff if I can help it.
I loved the old version of Hyphoria, I'm loving this one! Thanks <3
@Crix_69 Thanks for the tip! Glad you enjoy the new one as well, I'm quite proud of it.
Ever since I've discovered your V0.02 model I can't stop using by default. The prompt adherence is one of my favorite things. It likes to default to a version of the Kubrick Stare that I rather like.
It does get harsh at higher CFGs ( > 4.5 in my experience starts over-fitting). But if you've got a look/style in mind you want to prompt for already, its an excellent go to. Keep doing what you're doing.
Consider giving a 3 sentence primer on what the model is for