Used a different training strategy than my previous Illustrious models. SilvermoonMix-Illustrious-Evolved is good at:
Stable hands
Neutral art style to work with artist tags
Multiple characters
For instructions on how to use the V-pred model, please refer to https://civarchive.com/models/833294/noobai-xl-nai-xl
Recommended settings:
Sampling method: Euler a
CFG: 4-5
Start with no negative prompt. Sample images were generated without negative prompt.
Use ADetailer extension to enhance facial details: https://github.com/Bing-su/adetailer
推荐设置：
采样方法：Euler a
CFG: 4-5
先尝试不使用负面提示词。示例图片都没有使用负面提示词。
使用 ADetailer 扩展来增强面部细节：https://github.com/Bing-su/adetailer
Change notes:
v6.0: Further enhanced lighting, prompt obedience and compatibility with LoRAs that are trained on other models.
v5.0: Lights looks more nature now. Improved logic.
v4.0: Improved compatibility with LoRAs that are trained on other models.
v3.1: Improved logic, especially background. Adjusted saturation.
Description
FAQ
Comments (15)
May I know when to update the approximate dataset?
Hi @NaGaRi, this is a refinement of Noob VPred V1.0, specifically optimized for logical consistency and stylistic stability. This version does not introduce new concepts or characters. The knowledge base will only be updated once a model with newer knowledge becomes available.
@silvermoong thank you!
@silvermoong Have you tried implementing Chenkin Noob as that is the most up to date version of noobai?
@HaloSkull I have been thinking of this. Checkin Noob is likely to have another update in February.
@silvermoong Yeah the 0.3/1.0 version is supposed to release in February. From my usage, it seems to be more stable than noobai and has superior hands. This could be because no e261 data was added to mess up the danbooru dataset.
@HaloSkull The 'evolved' training strategy takes me lots of time and computing power, so I’ll probably wait for a stable version.
In the meantime, I'm thinking of training one using the traditional 'SilvermoonMix' strategy for a test.
嗯，大佬，可以问一下，你那张芙兰朵露的画风lora吗
@2912047909317 这个模型下的例图吗？这些都是模型直出的，没有使用画风lora
What's the adetailer settings?
@Dineko Hi, you may right-click one of my demo images, open it in a new tab and download. It contains the generation metadata. I believe WebUI can load it via the PNG info tab.
@silvermoong thanks mind share this
99coins_anime_girl_face_m_seg.pt
Best model for styles and artist tags I've ever used by far. It even does good with some anime screenshot character Loras (it tends to use the same angles sometimes though, and eyes are a bit unstable)
@justaguy111 Thanks so much for the kind words! Glad to hear you’re enjoying the model.