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..
Feedback Guidance of Diffusion Models
Authors: Felix Koulischer, Florian Handke, Johannes Deleu, Thomas Demeester, Luca Ambrogioni
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our novel state-dependent dynamic guidance scheme on Image Net512 512 using EDM2 models, where it consistently outperforms CFG and remains competitive with LIG, while benefitting from a strong mathematical framework. To assess generality, we additionally compute FIDs on MS-COCO in the T2I setting. Alongside these quantitative results, we provide qualitative examples showing that the self-regulated feedback guidance scale naturally increases with prompt complexity. |
| Researcher Affiliation | Academia | 1 Ghent University imec 2 Donders Institute for Brain Cognition and Behaviour, Radboud University |
| Pseudocode | Yes | Algorithm 1 Feedback Guidance (FBG) Input: Pre-trained conditional and unconditional network with prediction µθ(xt|c) and µθ(xt), mixing factor π, the two timestep hyperparameters t0 and t1 and a maximal guidance scale value λmax Derive δ, τ from π, t0, t1 and pmin from λmax Set hyperparameters (App. C.2) x T N(0, I) Initialize state log p(c|x T ) = 0 Initialize posterior and guidance scale λT (x T ) = p(c|x T ) p(c|x T ) (1 π) for t = T, . . . , 1 do µθ,guid(xt|c) = µθ(xt) + λt(xt) µθ(xt|c) µθ(xt) Compute and mix scores xt 1 = µθ,guid(xt|c) + σt 1|tz with z N(0, I) DDPM Step log p(c|xt 1) = log p(c|xt) Update the log posterior + τ 2σ2 t 1|t xt 1 µθ(xt|c) 2 xt 1 µθ(xt) 2 δ log p(c|xt 1) = max log p(c|xt 1), log pmin Clamp the posterior λt(xt 1) = p(c|xt 1) 1 p(c|xt 1) Update the guidance scale end for |
| Open Source Code | Yes | Our code is available at this link. ... The code is further provided at: https://github.com/Felix Koulischer/Feedback-Guidance-of-Diffusion-Models |
| Open Datasets | Yes | For class-conditional experiments, we use EDM2-XS Karras et al. [2024b,a] trained on Image Net512 512 Deng et al. [2009] with 64 function evaluations. ... To assess our approach beyond visual inspection, we compute FID, FDDino V2 and Aesthetic-Score6 on 3k MS-COCO prompts Lin et al. [2014], Heusel et al. [2017], Oquab et al. [2024]. |
| Dataset Splits | No | The paper mentions evaluating on 3k MS-COCO prompts and refers to standard benchmarks like ImageNet, but does not explicitly detail the training/test/validation splits used for their experiments. For example, it mentions "3k MS-COCO prompts" but not how they are split for evaluation. |
| Hardware Specification | Yes | For the stochastic sampler all experiments are run using on a NVIDIA Tesla V100-SXM3-32GB GPU. ... For the 2nd-order Heun sampler of the PFODE Karras et al. [2022, 2024b] a NVIDIA Ge Force RTX 4090 GPU is used. ... For the T2I results using Stable diffusion 2Rombach et al. [2022], we rely on an NVIDIA Tesla V100-SXM3-32GB GPU. |
| Software Dependencies | No | The paper mentions using EDM2 models and Stable Diffusion 2, but does not provide specific version numbers for these or any other software dependencies like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | For Classifier-Free Guidance, we search over the guidance scale λ. For Limited Interval Guidance, we optimize jointly over the guidance scale and the starting point σmax. For Feedback Guidance, we sweep t0, t0 + t1, and π. Optimal hyperparameters are listed in Table 3 and vary across metrics (FID vs. FDDino V2) and samplers. |