Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..

Token Perturbation Guidance for Diffusion Models

Authors: Javad Rajabi, Soroush Mehraban, Seyedmorteza Sadat, Babak Taati

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on SDXL and Stable Diffusion 2.1 show that TPG achieves nearly a 2 improvement in FID for unconditional generation over the SDXL baseline, while closely matching CFG in prompt alignment. These results establish TPG as a general, condition-agnostic guidance method that brings CFG-like benefits to a broader class of diffusion models. The code is available at https://github.com/Taati Team/Token-Perturbation-Guidance
Researcher Affiliation Academia Javad Rajabi1,2 Soroush Mehraban1,2,3 Seyedmorteza Sadat4 Babak Taati1,2,3 1University of Toronto 2Vector Institute for Artificial Intelligence 3KITE Research Institute 4ETH Zürich EMAIL EMAIL EMAIL
Pseudocode Yes Algorithm 1 Token Perturbation Guidance (TPG) for Diffusion Models Require: Noisy input x T N(0,I), shuffling matrices Sk,t, set of perturbed layers L, score function (denoiser) sθ, guidance scale γ, total time steps T 1: for t = T, . . . , 1 do 2: // Forward pass without perturbation 3: s+ θ (xt) sθ(xt, t) 4: // Forward pass with perturbation 5: Run the network sθ(xt) a second time with the following modification to have s θ (xt): 6: for each layer k do 7: if k L then 8: Apply token perturbation: H k Sk,t H k 9: // Apply token perturbation guidance 10: sθ(xt, t) s+ θ (xt) + γ(s+ θ (xt) s θ (xt)) 11: // Update sample 12: xt 1 Solver Step(xt, sθ(xt, t)) 13: return x0
Open Source Code Yes The code is available at https://github.com/Taati Team/Token-Perturbation-Guidance
Open Datasets Yes Moreover, all experiments are conducted by generating 30k samples for each method (unless stated otherwise) and evaluated on the MS-COCO 2014 validation set [32].
Dataset Splits Yes Moreover, all experiments are conducted by generating 30k samples for each method (unless stated otherwise) and evaluated on the MS-COCO 2014 validation set [32].
Hardware Specification No No specific hardware details (GPU models, CPU types, memory) are provided within the paper text.
Software Dependencies No No specific software versions (e.g., Python 3.8, PyTorch 1.9, CUDA 11.1) are mentioned in the paper.
Experiment Setup Yes In all experiments, the compared methods are evaluated using their original configurations and default guidance scales. The guidance scale for TPG is fixed at 3.0.