Looped-DiT retrofit of FLUX 3 Action 7B (AgiBot World Beta video prediction)
Looped Diffusion Transformer (Chng et al., 2026, arXiv:2609.40305) retrofit of
Jinyan0924/flux3-agibot-videoft (itself a full fine-tune of black-forest-labs/flux-3-action-base): the 33
joint blocks are split 12 / 9 / 12, the middle 9 are run N = 4 times with head-wise gated attention, and every
exit is trained with deep supervision. Task: k past head-camera frames + caption → 33 future frames at 15 fps,
384×512.
Files
| file | what |
|---|---|
ema_loop2500_step-2500.safetensors | r1: 2500 steps (LR 1e-5, 500-step cooldown), EMA 0.999, bf16 state dict |
ema_loop4500_step-4500.safetensors | r2: resumed from r1's step-2500 state and trained to step 4500 at constant LR 1e-5 (no cooldown; the planned 5000-step run with cooldown was stopped early after repeated infrastructure crashes), EMA 0.999, bf16 |
checkpoints/step-4500/ | full r2 checkpoint (FSDP2 DCP model + optimizer shards, CPU EMA shards, rng, trainer state) for resuming |
r1/checkpoints/step-2500/ | full r1 checkpoint |
metrics.jsonl, eval/val.jsonl | r2 training curve and periodic val losses (EMA weights, fixed windows / timesteps / noise) |
random_depth_n1-8/ | side run: same retrofit trained with a random loop depth N ∈ {1..8} per step (2000 steps) |
Val loss (EMA, N = 4, lower is better; k = number of context frames): r1 step 2500 k1 0.0787 / k9 0.0654;
r2 step 4000 k1 0.0783 / k9 0.0651 (the step-4500 val pass was interrupted by the crash that ended the run, so the
last val point is step 4000). Test-split PSNR / SSIM / LPIPS / FVD and the loop-depth sweep (N = 1, 2, 4, 8 at
inference) are in agibot-flux3/results/RESULTS.md of the GitHub repo below. Training finished 2026-10-16.
Load an EMA file with the ldvflux package from the GitHub repo (ldvflux.eval_phase4 shows the full
sampling path); the weights are a plain safetensors state dict of the looped DiT.
Code, configs and results: https://github.com/naomili0924/looped-dit-video/tree/main/agibot-flux3
License and use
Non-commercial use only. The weights are a Derivative of the FLUX Model and are distributed under the
FLUX Kommunity License (LICENSE.md, attribution in NOTICE); the training data, AgiBot World Beta, is
CC BY-NC-SA 4.0. No dataset frames or captions are included here.