Attention: See version details for a full list of trained artists, the list is too big for the example images alone. Also, this model was trained using v pred like loss and will fry at very low CFG. I recommend using the CFG rescale extension with it.
Version 5 is out. It is likely the last version of olympus in 1.5. If I make another one, I will likely do it in XL. In this version I increased the dataset to ~28000 curated images, trained from scratch for about 60 epochs on a cosine learning rate schedule. For best results, be thorough in your prompts. This model follows what you prompt very closely, which is good and bad. As a general rule, if you want something to appear, say it in the prompt and don't assume it will just happen. I have taken care that many uncommon tags in most models (such as pubic hair and upside-down), including many hardcore tags, are represented and working. See the example images for examples, they are straight text to image without any inpainting or adetailer. Also, make sure to use some of the 80 different artist styles available to the model. These were all trained in this version and don't require loras. That is the best way to achieve great gens with this model.
This is my first attempt at fine tuning something. I focused on fine tuning styles and concepts that I can't easily find in a versatile base model out there,as well as providing good hentai capabilities. Model was trained on ~9k images for ~150k steps, Clip Skip 1. See example images for example concepts and styles that I finetuned in. The model is very versatile and still surprises even me. It does very well with both women and men, even though the latter does not show up too much in the example images (sorry folks, there's just too much to showcase in only 20 images). It also knows a surprising amount of characters, I encourage everyone to try out a few and see what you get. All images were generated without adetailer or inpainting of any kind, using only hires fix with 0.3-4 denoise.
Description
27 new trained artists. Full list:
Ishikei
satsuki neko
cutesexyrobutts
dino (dinoartforame)
calm (artist)
krystopher decker
nababa
nat the lich
gogalking
hungry clicker
takayama toshiaki
kase daiki
sage joh
masami chie
bluethebone
urushihara satoshi
optionaltypo
paloma piquet
iuui
xilmo
stanley lau
uo denim
anato finnstark
hews
ratatatat74
ataruman
bartolomeobari
Training setup:
This model was trained for 100 epochs using the cosine learn rate scheduler in kohya. base learning rate was 4e-6. Some relevant configs:
```
"max_resolution": "640,640",
"min_snr_gamma": 0,
"noise_offset": 0.1,
"v_pred_like_loss": 0.75,
"batch_size": 6
```
FAQ
Comments (14)
Interesting model for sure. Some unique gens.
your models are legitimately the most responsive to prompts than any model I've ever used and I've tried hundreds. It also doesn't have "same face" syndrome which is refreshing. Absolute stellar work!
Thanks, it's great to hear people like it.
Really well done on the model. Do you happen to have a training guide for checkpoints? There's a ton of resources explaining how to make loras but almost nothing on trained checkpoints. I would really appreciate it if you could share the settings/tools you used. Thanks.
Sure, I'll try to write something in the main model description, look for it later (I am busy atm). But the real help here, I guess, is having a diverse and balanced dataset.
@rocp Hey just following up. If you get the time, could you please explain how you trained this model?
@screamlouder Hey, I added some notes to the version info. It's not much, but I don't know what else to write. I usually hang around the unstable diffusion discord server. It's easier to talk over discord, so if you're ever in that vicinity, ask around for me and we can talk.
How to fix muted colors?
Hard to say without an example, but that usually happens because you are lacking a VAE. If you click on files for version 1.0, you can download a vae file that I recommend. Put it into your models/vae folder, then select it in settings.
Hi! first of all amazing model! I wanted to know what upscaler is recommended and what resolution i should generate with?
may I ask what VAE i should use?
and also how many hires steps
@melonnarian344
A lot of questions, lol.
I use vae-ft-mse-840000-ema-pruned as VAE, you can use whatever you like, but this one is likely the better one.
The model was trained with bucketing, so it should be pretty flexible with generation resolution. That said, there's more data in 512x768 and 768x512 than in the other buckets, so those resolutions will likely work better.
For the upscaler, I've been using Valar 4x from the upscalers wiki (look it up), but that's again just personal preference. A lot of people love the anime specific ESRGANs, so whatever works for you. I always leave my hires steps set to 0 to run the same amount as non-hires steps.
I like the SDE Karras sampler at 20 steps or 2M SDE Karras sampler at 50 steps. You can experiment with others and see what you get.
@rocp haha sorry for the inconvenience but thanks a lot man
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.



















