Raehoshi Illust XL
an enhanced iteration built upon the Illustrious XL model. It aims to elevate the visual style by addressing some of the limitations in the original, such as oversaturation and artifact noise. While these issues are not entirely eliminated, noticeable improvements have been made. The goal is to deliver a more polished, balanced output while staying true to the strengths of the base model.
Why Early Access?
Early access helps keep the project going. I don’t have my own GPU, so all training is done through rented cloud GPUs and that gets pretty expensive. By getting early access, you’re directly supporting the development of my models and helping me keep improving them. If you'd like to support me further, you can also buy me a coffee on Ko-fi! Every bit of help means a lot and keeps the future updates coming.
Recommended setting
Positive prompt :
masterpiece, best quality, very aesthetic, absurdresNegative prompt :
bad quality, worst quality, jpeg artifacts, sketch, bad anatomy, signature, watermarkSteps : 25+
CFG : 5-7
Sampler : euler a or dpm++2m karras (euler for vpred)
Standard resolution :
832 x 1216, 1216 x 832, 1152 x 896, 896 x 1152, 1344 x 768, 768 x 1344, 1024 x 1024High resolution :
1024 x 1536, 896 x 1536, 1536 x 1024, 1536 x 896Hires.fix Setting:
Upscaler : 4x Foolhardy Remacri
Hires step : 10-15
Denoise : 0.1-0.3
Special Tags
Quality Tags:
masterpiece
best quality
good quality
average quality
bad quality
worst quality
Rating Tags:
safe
sensitive
nsfw
nsfw, explicit
Aesthetic Tags:
very aesthetic
aesthetic
displeasing
very displeasing
Training Details
The model was developed using a two-stage fine-tuning process. In Stage 1, new series and characters were introduced into the model. Stage 2 focused on fixing issues and enhancing the overall style for improved output.
Stage 1
Dataset : v1-31k, v2-37k, v3-34k, v4-60k, v5_v5.1-18k, v6-15k, v7-39k, v8-41k, v9-30k, v10-30k with multi resolution
Hardware : 2x A100 80gb, v3, v4, v5, v5.1-2x H100 80gb, v7,v8, v9, v10-RTX PRO 6000
Batch size : 32
Gradient accumulation steps : 2
Learning rate : 6e-6
Text encoder : 3e-6
Epoch : 15
Stage 2
Dataset : v1-2.5k, v2 and v3-2.3k, v4-2.5k, v5-2k, v5.1-1.8k, v6-1.5k, v7-1.7k, v7.1,v8-4.1k, v9-1.9k, v10-2.4k
Hardware : 1x A100 80gb, v7_v7.1,v8, v9, v10-RTX PRO 6000
Batch size : 48
Gradient accumulation steps : 1
Learning rate : 3e-6, v5.1-2.5e-6
Text encoder : disable
Epoch : 15
List of New Series/Characters Trained:
Zenless Zone Zero
Wuthering Waves
Honkai: Star Rail
Genshin Impact
Arknights: Endfield
Umamusume
Azur Lane
Arknights
Fate/GO
Dandadan
Make heroine ga oo sugiru
Kusuriya no Hotorigoto
Hololive from justice and dev is
Indie Vtuber Dooby, Yuuki Sakuna, Nimi Nightmare, and S***
100 girlfriends who really love you
Haite kudasai takamine-san
Alina clover
Nikke: bready and little mermaid
Kpop Demon Hunters
Full character list are available article here
For character trait details prompts, please refer to the Danbooru site for accurate tags and references.
License
Special thanks to Joe for supporting my works
Special thanks to Juno for supporting my works and help me with early tester
Description
Update knowledge same as v9 epsilon
Full Character List
Wildcards
FAQ
Comments (19)
more vpred
what is the difference between v-pred 2.0 and rahoshi 9.1?
@Rangiku209090 Raehoshi v9.1 uses epsilon, while vpred v2.0 uses vpred. Both versions have different base models. However, the vpred version has limited WebUI support, it does not work with A1111 and requires Forge/ReForge or ComfyUI instead
bro crazy how i commented this just before vpred 2.0 was released
@Raelina May I know how do you make the checkpoint ? are doing it locally?
@Rangiku209090 Actually, I don't train them locally. As mentioned in the 'Why early access?' section of the model description, all my models are trained by rent a cloud GPU such as RunPod. Full finetuning requires high-end GPU and large VRAM (you can see the specific GPU models used in the 'Training Details'). I only use my local machine for generating the image showcases
I haven't been able to test 9.1 as much as I wanted, but just wanted to provide some feedback.
Its a big improvement over 9.0, but I noticed some characters with specific outfits are hard to reproduce, mostly characters from 7.0 ish,
yixuan from ZZZ is one example, since her sleeves are quite specific.
If I'm able to do more testing, I will let you know. Either way, great work on the model update, definitely was an improvement.
Thank you for the feedback. If you find any other characters, please let me know.
@Raelina After doing some more testing in my free time, I think the only characters affected are the ones with complex or unique outfits. Most of the details stay, but the difference I'm noticing mostly is that newer character's details are like 95% accurate, while older characters accuracy is like 80~85% which is as expected. I did an experiment and I merged 7.1 with 9.1 and noticed an improvement in general on the older characters. Dunno if this would be useful for your next model but wanted to share my discovery. I would recommend doing some testing on your side, it would be interesting to see what you find.
Lastly, dunno if I can make some characters/series suggestions to add?
I was thinking adding Overwatch new characters and Pokemon new characters (only the humans aka trainers and Gym Leaders) would be good additions. Since those are popular series, but I can see an issue with OW characters, since they tend to do fanart of all outfits.
Well, I hope my testing helps you a bit, I will look forward to your next itteration!
Are there any plans to update the noobeps version?
I'm still undecided about updating the noob-eps version, but I'll think about it. Thank you
Thank you so much for your work on this checkpoint. I appreciate creators who favor trained checkpoints over merges, as they truly understand how they work, unlike merges which, most of the time, become black boxes, especially when it comes to tags understanding.
Speaking of tags, i have a quick question. Tags like "masterpiece" and "best quality" come from the original Illustrious checkpoint training. So, "masterpiece" is meant to define images that 100% of the ratings agree are superb, while "best quality" is meant to define images that 80% (or 90%, i can't remember what the Illustrious release PDF says) of the ratings agree are superb. Therefore, what is the logic behind using both "masterpiece" and "best quality" in the prompt quality tag sequence?
At first glance, this seems illogical, and only one of these quality tags should be used. In this case, if we want to maximize quality, we should only use "masterpiece" since it only includes images that everyone considers superb, right?
Thank you for the kind words. You are technically correct, logically "masterpiece" should be enough since it's the highest rank.
However, in the world of diffusion models, more reinforcement usually leads to better stability. While "masterpiece" represents the absolute top tier, the dataset for "best quality" is much larger. By using both, the model can capture the elite details from "masterpiece" while maintaining the stability and variety of the "best quality" dataset. It’s basically a way to reinforce the signal and ensure the most consistent results
Very very good model. Looking for a noobai model which contains new characters from games and animes and this is just what I'm looking for.
Hello, author, which dataset is the most recent, your version 9.1 or vpred v2.0? I noticed that vpred v2.0 was updated later than version 9.1
Hello! Both versions actually use the same dataset, even though vpred v2.0 was released later. My workflow is to focus on fixing and finalizing the Epsilon version first. Once that's done, I move those updates over to the Vpred version. That’s why Vpred always follows slightly after the Epsilon release. Thanks for asking
@Raelina Okay, thank you for the explanation, author. Your large model understands more about characters than those of other authors, which is very good
Don't use skip_early_cond with this model in a1111 or reforge webui. It makes monochrome and greyscale sketch.
It didn't work; setting skip_early_cond to 0 still resulted in a grayed-out image.