help wanted
Repository metrics
- Stars
- (7,198 stars)
- PR merge metrics
- (Avg merge 18h 39m) (16 merged PRs in 30d)
Description
It's just my one-sided doubts, about the implement of the v-prediction. In sdxl training, the source code implements v-prediction by:
def add_v_prediction_like_loss(loss, timesteps, noise_scheduler, v_pred_like_loss):
scale = get_snr_scale(timesteps, noise_scheduler)
# print(f"add v-prediction like loss: {v_pred_like_loss}, scale: {scale}, loss: {loss}, time: {timesteps}")
loss = loss + loss / scale * v_pred_like_loss
return loss
which is mathematically equivalent to: L:=L+snr*L*w, where w=v_pred_like_loss, and snr=scale, while the paper suggests: L:=snr*L.
So, is the source adds additional v-pred like loss rather than scaling it? Why are the implementation and paper different? I'm not a mathematician, and maybe I'm short-sighted. Hope someone can answer my doubts :D