CoolFace
Modelpublic

armanakbari4/g1_fdmV2_allTasksLCM_1000

sourceHugging Faceapache-2.0updated 4mo agoView on Hugging Face
0likes
Model Card

g1fdmV2allTasksLCM_1000 — LingBot-VA G1 LCM-distilled transformer (step 1000)

LCM video-only consistency distillation of the joint 5-task G1 alltasks teacher (armanakbari4/g1_fdmV2_allTasks_7500 — step 7500 of the 10000-step FDM-v2 post-training run on JingwuLuo/all_tasks_lerobot). Target: 2-step video generation.

This repo ships the target student (EMA-frozen — the standard LCM eval target). The online student exists in our training output but is not uploaded; ask if you want it for comparison.

  • —Teacher: armanakbari4/g1_fdmV2_allTasks_7500 (transformer/)
  • —Recipe (distill_video_v2/config_g1_alltasks.py):
  • —distill_mode: video (video-only consistency loss)
  • —num_ddim_timesteps=2 (k=500 stride → target 2-step generation)
  • —lcm_skip_k=6, EMA decay 0.995, huber loss (huber_c=0.001)
  • —Teacher CFG range [2.0, 10.0]
  • —lr=5e-6, grad_accum=8, batch=1, 4×H100
  • —Optimizer step 1000 of a 2000-step run (only 500/1000 were saved before the run was stopped at step ~1107; 1500/2000 were never reached).
  • —This repo contains only the `transformer/` (LCM-distilled, EMA target) — vae/, text_encoder/, tokenizer/ are unchanged from robbyant/lingbot-va-base.

Tasks covered (instruction strings used during teacher training)

sluginstruction
open_lid_add_potatoOpen the pot's lid and put the potato inside the pot.
pick_red_bottlePick up the red bottle
pick_and_move_bottlePick the pink object and put it on the cross mark.
put_carrot_n_cupPick up the carrot and put it inside the blue cup, then put the cup on the cross mark.
put_cup_n_broccoliPick the pink object and put it in the orange basket, then pick up the broccoli and put it inside the pink object.

Assemble an eval-ready checkpoint

bash
hf download robbyant/lingbot-va-base                  --local-dir lingbot-va-base
hf download armanakbari4/g1_fdmV2_allTasksLCM_1000       --local-dir alltaskslcm_1000_dl

mkdir -p g1_alltasksLCM_1000
ln -sf $(realpath alltaskslcm_1000_dl/transformer) g1_alltasksLCM_1000/transformer
ln -sf $(realpath lingbot-va-base/vae)              g1_alltasksLCM_1000/vae
ln -sf $(realpath lingbot-va-base/text_encoder)     g1_alltasksLCM_1000/text_encoder
ln -sf $(realpath lingbot-va-base/tokenizer)        g1_alltasksLCM_1000/tokenizer

Serve with CONFIG_NAME=g1_alltasks MODEL_PATH=g1_alltasksLCM_1000 and set num_inference_steps=2 (the distillation target). transformer/config.json has attn_mode: torch (inference-ready).