-
Update figA_iwa_learn_vs_frozen_a0667 (full run9 curve)
-
Update appendix training-dynamics figures (GroundFlow only, SD-RPN removed) + nips.tex + plot script
-
Update doc/nips.tex (appendix training-dynamics figure) + add visualize plot scripts
-
Add appendix training-dynamics figures (figA_loss_{4b,8b}, figA_gn_{4b,8b}, figA_iwa_learn_vs_frozen_a0667)
-
Add 32B SD-RPN TensorBoard events under training_logs/32b_sdrpn/
-
Add 32B SD-RPN checkpoint (qwen3vl-32b-dense-roi-K49T3-150k-sdrpn, step 1173, train_loss=0.0226)
-
Shorter trapezoid sides for source_grid_oblique
-
source_grid_oblique: shorter sides (flatter, 720x280)
-
Update source-grid trapezoid corners + height (flatter)
-
Flatter lying-down perspective for source_grid_oblique
-
textvqa patch tiles: switch to camera image; ROI + SINK + attention figures (PNG)
-
32B full results: nips.tex (Vanilla/SD-RPN/GroundFlow), trainer_state for run2 + SD-RPN, textvqa patch tile figures (PNG), eval/sync scripts
-
Upload GroundFlow training logs (4B/8B/32B) + refresh main results table
-
sync visualize from /opt/tiger/thothvl_pretrain (HF tokens redacted)
-
sync figures from /opt/tiger/thothvl_pretrain (HF tokens redacted)
-
sync doc from /opt/tiger/thothvl_pretrain (HF tokens redacted)
-
sync QWENVL-PRIVATE from /opt/tiger/thothvl_pretrain (HF tokens redacted)
-
initial commit
