Online generation
DREAMBOX is a model based on illustriousXL_v01 and trained with over 17000 images. It utilizes Danbooru tags and natural language for labeling and includes over 100 trigger words. When using DREAMBOX , it is recommended to experiment with different combinations of trigger words to achieve various art styles.
DREAMBOX是基于 illustriousXL_v01为底模，并使用 17000 多张图片进行训练而成的，同时使用了danbooru tags和natural language进行打标并设定了超过 100 个触发词。在运用DREAMBOX时建议通过对不同触发词的排列组合，以获取不同的画风。
Recommended Settings:
Sampling Method: DPM++ 2M SDE
Schedule Type: SGM Uniform
CFG Scale: 6-4
建议设置:
采样方法：DPM++ 2M SDE
Schedule type：SGM Uniform
CFG Scale：6-4
Art Style Tag List:
- https://huggingface.co/rim0/GENESIS/blob/main/Art%20StyleV2.txt
Nji Style Tag List:
- https://huggingface.co/rim0/GENESIS/blob/main/Nji%20StyleV2.txt
Lighting Effect Tag List:
- https://huggingface.co/rim0/GENESIS/blob/main/Lighting%20Effect.txt
Character Tag List:
- https://huggingface.co/rim0/GENESIS/blob/main/CharacterV2.txt
Here are some combinations I use, feel free to try them out:
下面是几个我在使用的几个组合，欢迎各位尝试:
- (kuroduki (pieat)),(alzi xiaomi:1.1),(rei (sanbonzakura):1.1),(koujuan:1.1),nose,lips, (chromatic aberration:1.1),(colored inner hair), eyeshadow,(shiny),(shiny hair), multicolored hair,glowing eyes,
- (xilmo:1.3,)(nj (imai):1.2),(wenjun lin:0.9) (nj (aoai):1.1), 
- (wanke:1.2),(mon (monmon2133):0.9),(mochizuki kei:0.95), (tira 27:0.7),
- (akizero1510:1.25),(am314:1.15),(agoto:1.05),((bai qi-qsr:0.9),
- (tianliang duohe fangdongye:1.35),, (asagina gi:1.55),(kaneni:0.9),,
(zukky:0.75),((jazz jack:1.2)), (nep:0.9),(urakata (uracata):0.75),(tira 27:1.2),chibi,
- (alzi xiaomI:1.3),(mogumo:1.2),(fukuro daizi:1.2),(starshadowmagician),
- (esuthio:1.2), (k-suwabe:1.2),(nj (aicd):1.4),(john kafka:0.9), 
- (nj (aifh):1.25),(yaziri:1.1),nj (opai),(nj (aicd):1.1),(photorealistic),(game cg),(so-tora,cuso4 suiwabutu:0.9), 
- (jima:1.35),(ohisashiburi:0.9),(kupuru (hirumamiyuu):1.2) (fukuro daizi:1.15),
Description
Enhanced the model's responsiveness to natural language prompts.
Art Style Tag List:
- https://huggingface.co/rim0/GENESIS/blob/main/Art%20StyleV4.txt
Nji Style Tag List:
- https://huggingface.co/rim0/GENESIS/blob/main/niji%20StyleV4.txt
Lighting Effect Tag List:
- https://huggingface.co/rim0/GENESIS/blob/main/Lighting%20Effect.txt
Character Tag List:
- https://huggingface.co/rim0/GENESIS/blob/main/CharacterV3.txt
FAQ
Comments (12)
It's so impactful, I really like it
My personal top 3 in Civitai so far! Very unique texture in image!
Wow, top 3 is huge praise! Thank you,I'm really glad you like it.
Great model, but it's difficult to change art style to something from basic Illustrious model. It's the only big disadvantage of this model. It have great overall composition, but limited on number of styles for it. Not sure you have a time for it, but try to make an altered copies of images in your dataset - they should be in another art styles (filter or a moderate AI change). More different art styles - better (to not let to form a style bias). It may be something interesting. Model should learn to save only composition, not a specific style.
Hey, thanks for the thoughtful feedback and the great suggestion.
The model does have a strong stylistic bias, and that's largely intentional. The dataset is heavily curated with artists I personally like, which means it inevitably causes some 'knowledge overwriting' from the base model.
That being said, I'm always looking to expand its range. If there are specific artists or styles you feel are missing and would be a good fit for this model's direction, feel free to recommend them. I can't promise I'll add everything, but I'll definitely take a look when I work on the next version of the dataset.
@EWIGKEIT0 I like what it generates right now, so I don't think dataset really need something completely new. Indeed, there is a big room for further experiments, but I'm talking about current v4.0. It looks great, but it can be even better and there is almost 100% way to do it without a risk to break it. Some automation and whole dataset can be converted into something "unique". Imagine a model with the same "line art" as now, but it will generate color palette/traits of a specific artist in colorized version. It sounds great, I think.
This is not a question "what can you add?". This is question "what can you save from base model?". Base model have a lot of useful things and it's not a good decision to rewrite actually positive sides of it.
@Shio_N Hey, thanks again for the super detailed analysis. You've given me a lot to think about, and your insights are pretty much spot on.
I understand your point of view completely. Preserving the versatility and broad knowledge of a base model is a very valid goal for development.
However, my personal goal for this project is a bit different. I'm less focused on preserving the base model's original capabilities and more focused on training a model to perfect the specific art styles I personally enjoy.
That's also why I provide the Genesis LoRA separately. For users who prefer maximum flexibility and want to retain the base model's features, using the LoRA with adjustable weights is probably the ideal approach.
@EWIGKEIT0 I'm not sure why your answers look AI generated, but this is ok, I guess. If you have no intention to change direction then I should do it... I will try to mix it with other models block by block to try to save what I think is the best. Maybe something useful will appear. My PC can't handle training, but mixing is not a problem. I will add a link to your model in a case of a good result.
生成国风内容最正确的模型，还得是国人
The model seems unique and your gens are cool, but I can't get the eyes or face to look right. I'm not sure what I'm doing wrong. I'm using v4. Also there's a tendency for it to generate signatures.
You can try using higher resolution and Hires. Fix when generating images.
Regarding the issue of signatures, try adding negative: signature, artist name, copyright name, watermark, text, English text
Best Model BY FAR