Which version before v22 preserved real-world faces the best?
There are too many to test easily, could someone please provide some guidance?
v22 is uploading right now. Give it a moment and try that first. Keep in mind, there are lots of factors that may make an older version worse in other areas (which is why I always recommend using the latest).
v18.1 seemed to be well received, but it is a mixed bag with slight variations with each release.
v22 is uploading right now. Give it a moment and try that first. Keep in mind, there are lots of factors that may make an older version worse in other areas (which is why I always recommend using the latest).
v18.1 seemed to be well received, but it is a mixed bag with slight variations with each release.
Thank you for your reply. I already have v18.1. I'll compare it with v22 later. Thank you.
I tested V18.1/v19/V20/v22, and v18.1 performed best in maintaining facial consistency.
I tested V18.1/v19/V20/v22, and v18.1 performed best in maintaining facial consistency.
I'd really like to see an example of v18.1 having better facial consistency than v22 (with a workflow, just the faces cropped).
I believe that version updates don't necessarily mean that previous problems have been fixed; rather, different versions have different advantages and disadvantages. Therefore, I hold multiple versions simultaneously.
I believe that version updates don't necessarily mean that previous problems have been fixed; rather, different versions have different advantages and disadvantages. Therefore, I hold multiple versions simultaneously.
Yeah, this. I tweak stuff and upload stuff. I do run tests in attempts to fix issues, but AI models are squishy and it is hard to predict how all use cases will be affected. Generally, I'm pretty happy with the latest version until I make a new one 😁
版本太多，我觉得不要搞aio版本了，拆分，这样可以减少硬盘占用。
I believe that version updates don't necessarily mean that previous problems have been fixed; rather, different versions have different advantages and disadvantages. Therefore, I hold multiple versions simultaneously.
Yeah, this. I tweak stuff and upload stuff. I do run tests in attempts to fix issues, but AI models are squishy and it is hard to predict how all use cases will be affected. Generally, I'm pretty happy with the latest version until I make a new one 😁
Looking forward to your next update
I'm currently using version v20, which has the best balance and colors that are very close to the original image.
版本太多，我觉得不要搞aio版本了，拆分，这样可以减少硬盘占用。
You don't need to download and save them all. I actually only keep the latest one on my system.
版本太多，我觉得不要搞aio版本了，拆分，这样可以减少硬盘占用。
You don't need to download and save them all. I actually only keep the latest one on my system.
他意思應該是說 不需要提供VAE跟 CLIP這樣更新能少下6GB
我現在用Qwen2.5-VL-7B-Instruct-abliterated.Q8_0 配合 AIO的GGUF版本...
12GB的4070S AIO會載入過大
剛剛看了還是V18.1比較接近原本... 尤其是眼睛跟眼袋... V22眼妝更重
人物由些微側臉轉成完全正臉就跑掉了...
I wonder why male's p...s is not normaly recognized in newer versions of Qwen AiO, other stuff seems very good ...
The Facial details are boched by some loras that are merged with the AIO. Please test before merging.