RDBT | MOSS
A SFT LoRA from a personal all-in-one dataset, mainly for digital art and drawings, focusing on lighting, textures. Covered from wallpaper level digital art, to 5-min sketch
LoRA is trained against MOSS base (not SFT). Cover images are also from MOSS base. It works on SFT, but you might need to lower the strength.
Dataset contains ~2k images, handpicked by professional amateur artists. The criteria for evaluation are not visual appeal, but the drawing technique and the level of completion.
There is no fixed trigger words or style tags. All images are captioned by LLM. They are manually collected so there is no metadata. And I'm too lazy to classify them. (Most of the style can't be classified in just few words tbh. It's hard to describe art.). So I used high caption drop rate hoping the model can guide itself from it's high noise steps.
Versions:
(11/7/2026) v0.2:
Won't nuke turbo LoRA and styles anymore. We are in business.
v0.1 (serious bug, misaligned timesteps):
Same dataset as Anima v3.0. This is a test run.
EDIT: timesteps were misconfigured because I changed training tool from diffusion-pipe to ai-tookit, and was porting my patches.
Sharing merges using this model is not allowed. This "restriction" won't affect anyone. It's only aimed at those who steal others' models to sell.
Description
FAQ
Comments (24)
style loras seem to perform better in combination with Lodestone's 8step dmd lora (just use a cfg of 2 or higher though and negative prompts)
This is promising, boss. Unlike the other general purpose, larger models, this is one seems to have actually been trained on the dark arts of the booru bible. Besides your usual magic on stabilizing, I'm hoping you can get it to match Anima or IL when it comes to style fidelity. It already seems to have some knowledge there, so maybe it just needs a little nudge.
the authors seem to be working on training a model specifically for anime
I'm more confident with Moss as alternative to Anima (replacing it or not would be differebt matter) because it has Booru tags training. Krea 2 while very good for realistic and general digital art, it is very sucks at interpreting otherwise very common booru tag. It can't even done panties/bikini aside properly!
I thought it could be an Anima replacement... but its not. It dont knows a lot of "not so popular" Anime Characters which Anima can handle without any problems. And its not like its only a few Characters...there are a lot missing.
@scarlizz Yes, and I don't think this specific model will replace Anima when it comes to character knowledge (maybe the anime-specific one the devs mentioned could). But it's still huge to have a general purpose model with all the knowledge this one does. It's unheard of, as far as I know, since even the "uncensored" ones are only trained on very restricted datasets and concepts.
moss for base style/background and qwen 2.1 for editing.
Model doesn't have to remember thousands of characters if it knows how to transfer the character from a reference sheet to the background. And that's what the editing model can do.
I'm really hoping they release an anime (editing) model good at anime style/character transfer. Then we'll never need style/character LoRA.
@scarlizz just temporarily, since it's not a model especially foe anime
Just heard about this Model and Search for RDBT Moss and to no surprise, the Goat delivered
你好，我很喜欢你的作品！想请问你，现在在Krea 2、Qwen Image 2.1 等模型各有优势的情况下，你为什么会选择继续深挖 MOSS呀？你觉得 MOSS 最值得继续挖掘的地方是什么呢？是因为它现在已经有一些独特优势（我觉得提示词遵循、构图控制和微调潜力上应该比Anima优秀，但画面审美与动漫风格可能还不如？），还是你觉得它还有很多潜力没发挥出来？挺好奇你对它未来发展的看法。
如果要追求极致质量，避免 slop，格外风格和角色 LoRA 是必须的，即使基座模型有相关知识。所以动漫模型不需要记住成千上万的风格和角色（当然如果可以那就更好了，微调能快一点）。
有基本的动漫知识就能微调 LoRA。Krea2 是第一个有基本的动漫知识大尺寸模型。之前的模型，比如 F2 Klein，最基本的脸都不会画，再怎么微调都没用。
训练数据集上 MOSS 比 Krea2 更进一步，还支持 bounding box 让构图更可控。
@reakaakasky thank you
The black magic "stabilizer" still works. Got low quality output? Style LoRA overfitted and can't use strength>0.8? Add this LoRA. Boom, everything is fixed.
Works on every style LoRA I trained, downloaded, and even finetuned ckpt. E.g. https://civitai.red/models/3016505/mossad-demo-moss-image-danboorue621-57-artist-style-finetune
This comment was written when v0.1 was the latest.
OP, you're the only reason why I check my CivitAI notifications. I felt like this LoRA was a step in the wrong direction style-wise. Out of the previews, I liked all of the original comparisons more except for the armored warrior on horseback. I felt like all of the RDBT versions had subtle tells of AI that the MOSS counterparts didn't have.
My concerns are about the style, it's less soulful. This LoRA improves image quality and clarity at the noticeable cost of a different kind of aesthetics.
Thanks, I hope for an update in the future.
Thanks for the feedback.
Moss v0.1 used same dataset as Anima v3.0. It's not a wrong direction. It's just the old trick didn't work out well on MOSS. (see model page, why I use high caption drop rate.)
With Anima you can use @artist tags to guide the model to very specific styles. But with MOSS you can't. Without style guidance, the styles from my dataset are mixed up, and becomes a unified AI style (painting + drawing + digital art + ....).
I'm still waiting the anime version. And find a way to properly caption my dataset with fixed style trigger words (not likely).
edit: v0.1 still works as usual if you have strong style trigger words, like on this finetuned model, or with style LoRAs. https://civitai.red/images/146568174
It appears that the MOSS model has issues within the residual streams of certain channels, which causes it to perform worse when trained using bf16 compared to fp16 mixed precision.
Has been fixed now: https://huggingface.co/KBlueLeaf/MOSS-Image-heal
Back in days we have lumina and z-image with 10^15 residual streams, iirc. Because I recorded those models for calibrated fp8.
I just feel 80% claims in that AI generated report are very questionable.
I mean you are free to question Kohaku, but he is still "the" Kohaku, so...
nah, lower activation is always better. It just feels wired the AI considered so many aspects, except checking backward pass.
@reakaakasky I'm training with SDNQ uint8 and int8 matmul quantization in LoRa training. I wonder if it would make sense in this scenario to switch for the "healed" bf16 base.
@deitychaser I'm using bf16 with int8convt dit. adamw and lr0.00005. And for some reason I still think it learned too well (LoRA is overcooked)
But this LoRA is not for new concepts, it's just a tiny style fine-tuning. So I can't tell if the residual steam is big problem.
@reakaakasky the hood talk is that Kohaku made that "healed" base first and foremost for his own full booru tune, which he is preparing to conduct now. So either way he seems to put his money where his mouth is and considers this to be the better strategy for a very large finetune.
"MossRAEL" also learned "too" well and LR had to be adjusted down because of that after couple epoches https://civitai.red/models/3031912/mossrael-moss-anime-finetune
It’s a bit disappointing that MOSS’s knowledge regarding NSFW content might be even weaker than Anima’s. Even when using the official "skill" prompt alongside Gemini4a—and after numerous iterations—it often still doesn't perform as well as a simple string of tags used with Anima. That said, it’s genuinely impressive when it comes to creating manga, so I’m looking forward to an official "Anime" version that hopefully brings a broader knowledge base. (Also, I feel like its ability to replicate specific art styles is only at about 85%—maybe 90% if you’re lucky.A friend of mine jokingly referred to MOSS as an "oversized Anima," specifically in terms of its fidelity)