FACT: Failure-Aware Causal Training for World-Action Models
Paper • 2608.10232 • Published
import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("Bariona/fact-wam", dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]FACT — RoboTwin Checkpoint
Action-transformer checkpoint for FACT (Failure-Aware Causal Training for World-Action Models), fine-tuned from Wan2.2-TI2V-5B on the RoboTwin 2.0 benchmark (data).
Paper: FACT: Failure-Aware Causal Training for World-Action Models
Project page: https://fact-wam.github.io
Code: https://github.com/Bariona/FACT
Usage
The repo ships the matching normalization stats (norm_stats_delta.json), so inference/eval needs no dataset download or training:
huggingface-cli download Bariona/fact-wam --local-dir ./models/fact-wam
python -m scripts.inference_server \
--model_id ./models/Wan2.2-TI2V-5B-Diffusers \
--transformer_path ./models/fact-wam/transformer \
--stats_path ./models/fact-wam/norm_stats_delta.json \
--port 8093
For closed-loop RoboTwin evaluation, set TRANSFORMER_PATH=./models/fact-wam/transformer and STATS_PATH=./models/fact-wam/norm_stats_delta.json in evaluation/robotwin/launch_config.yml.
- Downloads last month
- -
Model tree for Bariona/fact-wam
Base model
Wan-AI/Wan2.2-TI2V-5B-Diffusers