R
ReflectIris

Reflect ML Dashboard

Live from 5 machines — reflect-ml-02

Mirror Metric v5Live
Live from 5 machines · Best: reflect-ml-02
Details →
68/100
Mouth
30/40
Identity
12/25
Perceptual
13/20
Face
14/15
Most Realistic Render
ml-01 — step 62,000
MM 79/100Train PSNR: 37.6 dBLPIPS: 0.202SSIM: 0.859200,000 Gaussians
exp 191 clean baseline
Realism PushMirror Metric 68/100

Looking like a real person

The avatar now preserves the patient's identity and produces convincing facial expressions. Mouth quality and perceptual realism are the current focus — the gap between "good" and "mirror-like" is where we're pushing hardest.

What's Working
  • +Best face reconstruction: 38.5 dB — the avatar looks like the person
  • +5 machines training simultaneously — testing multiple approaches in parallel
What's Next
  • Close the perceptual gap — making fine details indistinguishable from reality
  • Optimize for real-time rendering at 60fps on consumer hardware
  • Currently testing: cnn deform, clean baseline, expr opacity
Challenges We're Solving
  • We've run 231 experiments so far. Not all succeed, and that's by design — each one narrows down what works best
  • Real-time quality at 60fps requires balancing visual fidelity with rendering speed — we're solving both simultaneously
GPU Fleet — 5 Active
Live
MachineExperimentStepTrain PSNR / LPIPSMM
reflect-ml-02
exp 192 cnn deform
32% complete
63,150
32.9 dB Train
0.2028 LPIPS
68
reflect-ml-01
exp 191 clean baseline
33% complete
65,600
32.2 dB Train
0.2200 LPIPS
65
reflect-cart-02
exp 194 expr opacity
7% complete
13,750
31.1 dB Train
0.2531 LPIPS
59
reflect-cart-01
exp 193 mouth weight
21% complete
41,200
28.8 dB Train
0.2552 LPIPS
52
reflect-gpu
exp 195 blendshapes
12% complete
23,450
17.5 dB Train
0.4941 LPIPS
5