Best practice? #284
-
First of all, I would like to say thank you for the great work that has gone into roop unleashed. As a beginner, I wonder if there is some kind of best practice in how to proceed. I am often surprised by the results. In my opinion rather bad source images sometimes give better results than hq images. Is there a basic rule regarding the resolution of the source images? Is there a relationship between the resolution of the source image and that of the target? |
Beta Was this translation helpful? Give feedback.
Replies: 2 comments 1 reply
-
Yeah i've been trying to figure this out too (not related to resolution but similar). Some single images (not that high quality, like a cropped photo from the shoulders up not even 600px wide or tall) give me AMAZING results, consistently matching and sometimes even outperforming full facesets with multiple camera angles (and single images too obviously). Maybe there's a specific type of lighting that roop really loves?. Hopefully someone else with more experience can chime in and share their findings!. |
Beta Was this translation helpful? Give feedback.
-
Well, first of all roop and all of its forks and derivates are all using the same awesome swapper model from InsightFace. That model does all the groundwork of pasting the face from one to another and we are all treating it more or less like a blackbox. Now this model was originally released as open-source but when they developed better versions of it and started to create a business on it, they removed all traces of the original model and hide the better one behind a paywall. You can read about it here deepinsight/insightface#2294 |
Beta Was this translation helpful? Give feedback.
Well, first of all roop and all of its forks and derivates are all using the same awesome swapper model from InsightFace. That model does all the groundwork of pasting the face from one to another and we are all treating it more or less like a blackbox. Now this model was originally released as open-source but when they developed better versions of it and started to create a business on it, they removed all traces of the original model and hide the better one behind a paywall. You can read about it here deepinsight/insightface#2294
The model we are using does the face swapping at the low resolution of 128x128 pixels, the insightface paid model can do 512x512. That's why some authors inclu…