Working Kontext Face Replacement Process

Other Questions workflow
1

I’ve been experimenting with Kontext for face replacement and wanted to share a simple workflow that actually works pretty well. The basic idea is taking someone’s face from one photo and putting it onto another person in a different image.

The key thing I learned is that denoise settings really matter depending on your target image. If you have a close-up headshot where the face takes up most of the frame, you can push denoise up to around 0.95 and still get good results. But if you’re working with full body shots, you need to keep it lower like 0.90 or the face positioning gets messed up.

I’m thinking about adding some face cropping and upscaling steps to the process so I can use higher denoise values consistently. That should help get better facial resemblance overall.

Anyone else working with similar face swapping workflows? Would love to hear what tweaks or improvements you’ve found that work well. Also wondering if there are any good LoRAs out there that don’t completely change the chin shape - that seems to be a weak spot with the current Flux options I’ve tested.

2

Yeah, the chin shape thing is a real problem with most models right now. I fixed it by changing how I preprocess the source image - don’t crop tight around the face. Include more jaw and neck area so Kontext has better context for chin placement. For upscaling, do it after the face swap, not during. Keeps everything way more consistent and the facial features don’t get wonky. Don’t completely skip LoRAs though - try them at lower strengths like 0.6-0.7. They can actually help keep facial structure intact when you’re not going full strength.

3

i usually keep denoise around 0.88 too. seems to work nicely for me. adding feathering to your mask can really help with the blend, makes it look way more natural. good luck with your tweaks!