Radiance666/Progression-as-Latent-Drift

Progression as Latent Drift

Generative Forecasting of Slow-Evolving Pathologies (ECCV 2026)

Pretrained weights for Latent Drift, a two-stage progressive framework that forecasts patient-specific neurodegeneration by learning temporal change in a compressed semantic space rather than re-synthesizing full-resolution anatomy.

• 🌐 Project page: https://cutepkq.github.io/latent-drift/
• 💻 Code: https://github.com/CUTEPKQ/latent-drift
• 📄 Paper: https://arxiv.org/abs/2607.08270

About

Forecasting the future anatomy of slow-evolving neurodegenerative diseases could enable earlier intervention and better clinical-trial design, but it is hard because true progression signals are subtle in longitudinal MRI — across a typical one-year interval, disease-related change can be under 1% of the total volumetric variance. In this low-signal regime, transplanting modern generative sequence models fails in two structural ways:

• Identity collapse — the stationary background overwhelms the biological signal, so a model trained to synthesize the absolute future state collapses toward reproducing the current anatomy instead of learning the faint drift.
• Continuous interpolation trap — shifting the target to the temporal residual Δz still forces any Lipschitz-continuous predictor to interpolate the dense nuisance noise, smearing spurious change across the volume.

Latent Drift learns change in a compressed semantic space. A residual objective removes pixel-level identity from the prediction target (escaping identity collapse), and Finite Scalar Quantization (FSQ) acts as a topological dead-zone filter — noise below the quantization threshold is mapped exactly to zero while consistent structural drift is preserved; because rounding is non-Lipschitz, this breaks the continuous interpolation trap.

On longitudinal 3D brain MRI from ADNI + AIBL (3,981 current–future pairs), Latent Drift attains the best structural agreement (Diff-SSIM 0.8204, NCC 0.9880) and downstream clinical utility (Accuracy 88.33, F₁ 87.51), outperforming diffusion and autoregressive-transformer baselines.

Framework

Latent Drift predicts latent progression in two stages:

Stage	Component	What it does
1 · Latent Drift Tokenizer	LatentActionQuantization	Encodes the pre/post MRI pair, models the continuous latent drift Δz = zfut − zcur, and quantizes it into discrete drift tokens with FSQ
2 · Decoder-Transformer	Net2NetTransformer	Autoregressively forecasts the discrete drift tokens, conditioned on the patient's baseline anatomy and time gap, then decodes them back into the predicted future volume

Usage

# Download the weights
hf download Radiance666/Progression-as-Latent-Drift --local-dir checkpoints

# Run inference (see the code repo for setup)
python infer.py

Full training, data-preparation, and inference instructions are in the GitHub repository.

Citation

@inproceedings{feng2026latentdrift,
  title     = {Progression as Latent Drift: Generative Forecasting of Slow-Evolving Pathologies},
  author    = {Feng, Yuxiang and Wang, Juncheng and Xu, Chao and Hou, Wenlong and
               Wang, Huihan and Qian, Yijie and Liu, Yang and Sun, Baigui and
               Liu, Yong and Wang, Shujun},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

@article{feng2026latentdrift_arxiv,
  title   = {Progression as Latent Drift: Generative Forecasting of Slow-Evolving Pathologies},
  author  = {Feng, Yuxiang and Wang, Juncheng and Xu, Chao and Hou, Wenlong and
             Wang, Huihan and Qian, Yijie and Liu, Yang and Sun, Baigui and
             Liu, Yong and Wang, Shujun},
  journal = {arXiv preprint arXiv:2607.08270},
  year    = {2026}
}

License

Released under the MIT License.