You can click on the link below and try it out directly. If the effect is good, you can deploy it locally
https://www.runninghub.ai/ai-detail/2017076374867025921?inviteCode=rh-v1131
Fan benefits, register to get 1000 points, daily login 100 points, play 4090! Experience the super power of 48G!
prompt:
Preserve the subject's features and generate a high quality realistic human photograph
Description
FAQ
Comments (11)
How about making the opposite of this, like turn really image (cosplay) into anime
More realistic without using any lora. This is what, the 5th a2r lora that either doesn't do anything or makes the image less realistic than using no lora at all. It's so odd since Qwen cannot do it properly without lora, but with a2r in qwen it actually replaces the character with a real person, although textures are smooth and not good. Klein will make great textures without the need for any lora, but it simply won't replace the character with a real person, more so a stylized, smoothed out character reskined with better textures. All you have to do is pick any stylized character and prompt it to make it real, you can use any of these loras, you'll see their head will be huge and proportions will be wrong, something that doesn't happen as much with QIE (using a2r lora). I really hope someone can figure this out, otherwise this kind of lora, or prompting for making a real photograph, only work properly with already realistically proportioned drawings/semi-real
sad but true
wondering have you tried just prompting "completely realistic"? also there was a node to give more options for i2i
@Spamuellow What node is that? I tried so many types of prompt. I even went as far as prompting "remove the character and replace with a real person, stuff like that lol
It's spam, they are just trying to lure people to their runninghub ref/invite link
While it's possible to generate realistic results without LoRa, the aesthetics are often lacking. This is why we authors still train LoRa. I've uploaded a comparison image to the cover.
@zym0x RH is a platform that allows for quick results, surpassing Civitai in this aspect. Invitations allow authors to earn credits to train more models. I've spent a significant amount of time creating data and training results, and I'm sharing them for free; I don't see any problem with that.
@JayZ2015 Your examples were still realistically proportioned as the reference. Try it with cartoons or more stylized anime. You'll notice that even in your examples it made their heads actually bigger than the reference. I don't know why klein does that, but it sucks balls. Qwen will actually "remake" the characters head with a real person's head, the difference is night and day. It's not that this lora doesn't work, it does help retain characteristics (while no lora often loses a bit), but it's still not as effective as the same type of lora on Qwen for this task. And again, I'm 90% sure it's a model problem, I'm not sure loras will be able to fix that, unless someone makes a very strong "realism lora" that can help with that, without ruining the consistency.
@valkenweiss The issue of large head proportions does exist, and it's likely a problem with the training dataset itself. My dataset contains anime characters with realistic proportions, as I didn't specifically need to process chibi-style cartoon images. Perhaps other authors can achieve this.