Clean pixel arm: C3 + realism, EMA at step 60000
ema_060000.pt - the pixel-root arm of the clean line (outputs/clean/c3r_main/ema_060000.pt), trained from scratch as the
matched control of the voxel arm in 07_clean_voxel_f3r/.
- Matched to the voxel arm: same 3D-Belief trunk, same decoder warm start (free_completion_20260927 ray/step002000), same data (official RE10K train only, online VGGT-Omega teacher vggt_omega_1b_512, fresh 11-frame clip per visit), same realism losses (--drift-global --drift-geo --ot-p 0.5 --ot-mode lateral --ot-frac 0.3), seed 1, 8 H100 x 1 clip per step, constant lr (wsd, no cooldown). The only difference: no voxel head / fuser / voxel stream (pixel roots, coverage memory rule) - the five voxel flags of the voxel arm are absent.
- Stopped at 60k by decision (2026-10-28): still creeping (visible LPIPS W3 .241 -> .232 over 45-55k, .237 at 60k), far behind the voxel arm on observed metrics at matched steps. Not a converged plateau.
- Full command: command.json. Code: branch clean-re10k-online-teacher at f1252bf plus uncommitted changes to
flc/unified/train_v4.py (
--accumgradient accumulation, unused here at accum 1; redraw of clips with truncated PNGs): train_v4_vs_f1252bf.patch. - Held-out (40 clean RE10K episodes, unified 4, blind, T=1.0, coverage memory rule, no update head): eval_summary_40ep.json.
| Step 60k, no update head | Visible LPIPS W1 / W3 | Revisit-A PSNR W3 | Reveal LPIPS W2 | Whole KID | Patch KID occluded / out-of-view |
|---|---|---|---|---|---|
| pixel arm (this) | 0.263 / 0.237 | 23.59 dB | 0.303 | 0.0354 | 0.058 / 0.026 |
| voxel arm, same step (f3r_main 60k) | 0.214 / 0.206 | 25.10 dB | 0.207 | 0.0014 | 0.138 / 0.039 |
Pixel leads only on hidden-region patch KID. Update heads for both arms at 60k (upd.sbatch recipe) are pending.
- Load as an evaluate_stream_amodal.py --snapshot (dict with 'ema', 'step', 'args'), with --mem-rule coverage.
sha256 ema_060000.pt: e15fc5a7149802bbe7677ee6e9369b44e7f55cc54a28a4d2dc71faf366e1d820