CoolFace
Apppublic

WitneyWW/javisdit-edit-film-adalnsum-step12036

sourceHugging Facemitupdated 1mo agoView on Hugging Face
0likes
App README

JavisDiT edit — text_both_film_adalnsum_0827 @ epoch002-global_step12036

The FiLM adaln_sum variant (source summed into the timestep embedding, per-token adaLN) trained fresh from the base backbone, evaluated mid-training while job 489822 is still running. Both directions (remove, add) on the first 20 rows of the held-out test_first20_fixed.jsonl, each generated with and without the anchor frame — 80 generations, cfg 1.0, 50 sampling steps, per-entry seeding.

Both anchor modes are shown because this checkpoint sits mid-anneal: ref_prepend_warmup_steps: 30000 gives p_ref ≈ 0.60 at step 12036, so withref is the closest to in-distribution and noref is the deployment mode the anneal works toward.

Checkpoint frozen to _frozen_adalnsum_step12036/ (the run keeps only the 5 most recent epoch checkpoints; only ema.pt is read at inference).