FLUX.2 AI Clone Test: 65% Face Similarity, Then a Custom LoRA
FLUX.2 repaints the whole frame as one photo, so the light on the face finally matches the room.
· 21 views · YouTube video: The AI Clone That Got My Face 65% Right
Watch on YouTube · Subscribe to @PROOFOFDATA
Quick overview
- FLUX.2 repaints the whole frame as one photo, so the light on the face finally matches the room.
- Face similarity landed around 65%: recognizable, but not quite the real person.
- The fix in progress: a custom LoRA from 66 frames plus 66 face close-ups, about an hour on an RTX 4090.
- ~65% Face similarity with FLUX.2 out of the box
- 66 + 66 Full frames plus face close-ups used for LoRA training
- ~1 hour LoRA training time on an RTX 4090
What happened
Instead of pasting my face into a studio background, FLUX.2 generates the whole image as a single photo. That solves the classic AI avatar problem where the lighting on the face doesn't match the scene.
The trade-off was identity. Measured face similarity came out around 65%, close enough to recognize, not close enough to pass as me. So I set up LoRA training with AI Toolkit: 66 frames of me plus 66 close-ups of my face, roughly an hour on an RTX 4090.
Why it matters
Open-source models can now handle light and composition well, but a generic model doesn't know your face. A personal LoRA is the free route to teaching it what you actually look like, if you have a capable GPU and a clean set of photos.
The Short shows the setup. The full video covers every model I tried and the real cost of the whole experiment.
Watch the full test on YouTube: every model I tried and the real cost
FAQ
How long does it take to train a FLUX LoRA on your own face?
In this test, about an hour on an RTX 4090 using AI Toolkit, with 132 training images: 66 full frames plus 66 face close-ups. Times will vary with your GPU and settings.
Watch on YouTube · Subscribe to @PROOFOFDATA
Topics: ai clone, ai avatar, flux, flux 2, lora training, open source ai, face similarity, ai face


