The module already has VAE. Unless you have specific coloring requirements, it is recommended to use the VAE inside the module. Due to the unique nature of v-pred, I have created a dedicated section for v-pred.
(If the colors in version 2.0 are too heavy, you can switch the VAE to SDXL.)
Recommended Parameters:
Samplers: DDPM(suggest) > Euler > DDIM (⚠️ Other samplers will not work properly)
DDPM
CFG: 5.5 ( around 5.5)
Steps: 40 ( or higher)
Hi-res fix: Upscale 1.5x or higher, Denoising strength 0.15–0.45
Euler
CFG: 4~7
Steps: 20~35
Hi-res fix: Upscale 1.5x or higher, Denoising strength 0.15–0.45
Forge Configuration Guide (Update to the latest version):
Select SDXL from FreeU dropdown menu, set the start to 0.6 and end to 0.8.( V2.0 Not recommended SelfAttentionGuidance Integrated )
ComfyUI :
https://huggingface.co/Laxhar/noobai-XL-Vpred-0.5/blob/main/comfy_ui_workflow_sample.png
For Training v-pred LoRAs :
https://civarchive.com/articles/8723/vlora-or-how-to-train-a-lora-on-v-pred-sdxl-model
Adetailer :
https://github.com/Bing-su/adetailer
Description
V2.0
FAQ
Comments (9)
Woooow.... A new version. I actually remember you commenting here that it's "thankless and difficult job". Glad you still decided to do it, i've used this model for the longest at this point and it was my favorite(and only one that I used until recently) checkpoint for a very long time.
@ok_furry_image I don't think i've seen a single model that I tracked reach it's early access goal on this site though, but I get what you mean. Still - thanks for doing it, it's one of the best v-pred models on this site for sure and is really good even beyond bara and furry. Just not many people seeing this.
This Model is awesome! After setting up V-Prediction in SwarmUI (simply set "Override Prediction Type" to "V Prediction (ZSNR)") it performs better than other models
没想到vpred版本还会更新，但是相比1.0主要是提升了些什么呢？I didn’t expect that the vpred version would be updated, but what are the main improvements compared to 1.0?
画面更细致，背景更好，但稳定度有点变低
@ok_furry_image 感觉上这一版很少会出现shiny skin这种情况了，vpred给我的感觉依然还是比eps好用的多！希望继续更新！
如果你遇到草地，雪地等浅色大面积区域看起来过曝的问题，可以试试euler a采样器，cfg 4，step 30。虽然很多教程里推荐用cfg rescale功能减少过曝和过饱和，但我不推荐这个功能，因为会让颜色很黄，看起来像褪色照片。这一套设置在我的测试中是最合适的，天空，草地，水面，阴影，角色身上的轮廓光都没有出现色彩问题。另一个推荐euler a的原因是，不知道为什么在我的环境里，ddpm和euler a的结果几乎完全一致，但是euler a的细节更好一点。
v-pred的过曝，过饱和，色彩偏差的问题困扰我很久。我尝试了很多教程上的解决方法，下载了很多自定义节点，但没想到最后还是换采样器和cfg最有用。我曾经也尝试换回eps版本，但是体验过v版本的丰富色彩以后，再用eps版本就像蒙了一层灰雾。这个v-pred版本简直太棒了。
即使不更新，这个模型也已经非常出色。除非以后出现下一代模型。