TANGO

Test-Time Noise Guided Adaptation for Realistic Autoregressive Video Generation

Dimitrios Karageorgiou^1,2 · Symeon Papadopoulos¹ · Ioannis Kompatsiaris¹ · Efstratios Gavves²

¹ Information Technologies Institute, CERTH · ² University of Amsterdam

ECCV 2026

Inference

Any-length Video-to-Video Generation

python inference.py \
  --config_path configs/self_forcing_dmd_long.yaml \
  --output_folder "${output_dir}" \
  --checkpoint_path checkpoints/self_forcing_dmd.pt \
  --data_path "${videos_csv_path}" \
  --csv_root_dir "${videos_csv_root_dir}" \
  --use_ema \
  --num_output_frames ${num_output_latent_frames} \
  --num_input_latent_frames ${num_input_latent_frames} \
  --v2v

Autoregressive video diffusion models eventually collapse. Prior works aim to keep each frame on the manifold of real ones, but a trajectory whose every frame looks right can still reach a terminal point: a state in the manifold of real videos that the model lacks the knowledge to continue. TANGO detects terminal points at test time from the model's own noise predictions. Then, test-time adaptation is employed to steer the model away from terminal points, effectively trading inference-time compute for improved performance.

Citation

@inproceedings{karageorgiou2026tango,
  title     = {Test-Time Noise Guided Adaptation for Realistic Autoregressive Video Generation},
  author    = {Karageorgiou, Dimitrios and Papadopoulos, Symeon and Kompatsiaris, Ioannis and Gavves, Efstratios},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

Acknowledgments

Supported by the Horizon Europe projects ELIAS (grant no. 101120237) and ELLIOT (grant no. 101214398). Computational resources were granted with the support of GRNET.