CycleGAN: Cycle-Consistency Loss Live

Training Controls

Cycle-consistency weight λ 10

Learning rate 0.20

Training steps / frame 2

Live Metrics

Cycle-consistency loss ‖F(G(x))−x‖ + ‖G(F(y))−y‖

–

Discriminator accuracy (real vs fake)

–

Training steps

Point Clouds

Real X (sphere domain)

Real Y (torus domain)

G(X) — fake Y

F(Y) — fake X

F(G(X)) — reconstructed X

G(F(Y)) — reconstructed Y

How it works

CycleGAN learns two mappings, G: X→Y and F: Y→X, between two unpaired domains — here a sphere-shaped point cloud X and a torus-shaped point cloud Y. Two small discriminators D_Y and D_X try to tell real domain points from generated ones, and G/F are trained to fool them:

L_GAN(G,D_Y) = E_y[log D_Y(y)] + E_x[log(1 − D_Y(G(x)))]
L_GAN(F,D_X) = E_x[log D_X(x)] + E_y[log(1 − D_X(F(y)))]

With only the adversarial loss, G could map every x to a single realistic-looking y (mode collapse) with no relation to the input. CycleGAN fixes this with a cycle-consistency loss: mapping there and back should return you close to where you started.

L_cyc(G,F) = E_x‖F(G(x)) − x‖₁ + E_y‖G(F(y)) − y‖₁
L_total    = L_GAN(G,D_Y) + L_GAN(F,D_X) + λ·L_cyc(G,F)

λ slider — weight of the cycle term. Set it to 0 and watch the reconstructed clouds (light blue / light orange) drift away from their originals as G and F stop being inverses of each other — a live view of what cycle-consistency actually prevents.
Learning rate / steps per frame — control how fast the tiny 3→8→3 generator networks and 3→8→1 discriminator networks converge; both use real backpropagation computed every frame, not a scripted animation.
Reset weights — re-initialises all four networks from fresh random weights and restarts the step counter.

This is the same principle StyleGAN/CycleGAN-family models use for unpaired image-to-image translation (e.g. photo↔painting, horse↔zebra) — swap the 3D point clouds for pixel grids and the loss terms above are unchanged.

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.