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..
VarFlow: Proper Scoring-Rule Diffusion Distillation via Energy Matching
Authors: Huiyang Shao, Xin Xia, Yuxi Ren, XING WANG, Xuefeng Xiao
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we empirically validate the effectiveness of Var Flow. We conduct a comprehensive set of experiments across various standard benchmarks and diverse task settings. Refer to Section A for experiment setup and details. |
| Researcher Affiliation | Collaboration | Huiyang Shao1,2 Xin Xia2,* Yuxi Ren2 Xing Wang2 Xuefeng Xiao2, 1Tsinghua University 2Byte Dance Seed |
| Pseudocode | Yes | Algorithm 1 Var Flow Training Procedure Require: Student generator gϕ, data source qdata(x0), noise schedule ( αt, σt), energy score exponent β (0, 2), batch size K 2, latent distribution pz(z), time weighting w(t), learning rate η. 1: Initialize parameters ϕ for the student generator gϕ. 2: repeat 3: Sample a time t Unif[0, T] (or other suitable distribution). 4: Sample K latent vectors {z(k)}K k=1 iid pz(z), K real data points {x T,(k) 0 }K k=1 iid qdata(x0). 5: Generate clean student samples: x S,(k) 0 gϕ(z(k)). 6: Obtain noisy student samples x S,(k) t by applying the forward process (Equation (1)) to x S,(k) 0 . 7: Obtain noisy teacher samples x T,(k) t by applying the forward process (Equation (1)) to x T,(k) 0 . 8: Estimate the batch loss ˆLt(ϕ) using {x S,(k) t } and {x T,(k) t } per Equation (8). 9: Update student parameters: ϕ ϕ η ϕ ˆLt(ϕ). 10: until convergence 11: return Trained student generator gϕ. |
| Open Source Code | No | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [No] Justification: We will provide the source code and data after the paper is accpeted. |
| Open Datasets | Yes | Table 1: Image generation results on Image Net 64x64 (class-conditional), Image Net 256x256 (classconditional), CIFAR-10 32x32 (unconditional), and MS COCO 512x512 (text-to-image, zero-shot). Table 2: Quantitative comparison of state-of-the-art models across various architectures and steps for FID and CLIP scores on the COCO-10k dataset. References for datasets include: Deng et al. [2009] for Image Net; Krizhevsky and Hinton [2009] for CIFAR-10; Byeon et al. [2022], Schuhmann et al. [2022] for data used for MS COCO. |
| Dataset Splits | Yes | In this section, we empirically validate the effectiveness of Var Flow. We conduct a comprehensive set of experiments across various standard benchmarks and diverse task settings. Refer to Section A for experiment setup and details. Table 1: Image generation results on Image Net 64x64 (class-conditional), Image Net 256x256 (classconditional), CIFAR-10 32x32 (unconditional), and MS COCO 512x512 (text-to-image, zero-shot). Table 2: Quantitative comparison of state-of-the-art models across various architectures and steps for FID and CLIP scores on the COCO-10k dataset. (Implied standard splits for well-known benchmarks: ImageNet, CIFAR-10, MS COCO 10k). |
| Hardware Specification | No | Justification: We provide information about the computational resources (Sec. A) used for the experiments. (Note: Section A is not provided in the paper text). |
| Software Dependencies | No | Justification: We describe the algorithms and specific details (Sec. A) about the experimental setup, such as datasets used, specific hyperparameters (learning rates, batch sizes, parameters), training duration, evaluation metrics implementation, or baseline implementation details necessary to fully reproduce the claimed experimental results. (Note: Section A is not provided in the paper text, and no software versions are mentioned in the main body). |
| Experiment Setup | Yes | Algorithm 1 Var Flow Training Procedure Require: Student generator gϕ, data source qdata(x0), noise schedule ( αt, σt), energy score exponent β (0, 2), batch size K 2, latent distribution pz(z), time weighting w(t), learning rate η. Table 3: Ablation studies for Var Flow on SD1.5 (Lo RA) / MS COCO 10k (1-step). The "Optimal" configuration provides the baseline. Ablation Focus Configuration FID ( ) CLIP Score ( ) AES ( ) Var Flow (Optimal) β = 1, w(t) = σ2 t , Full Loss, K = 16 5.08 30.70 5.85 Energy Distance β β = 0.5 5.15 30.52 5.78 β = 1.0 (Optimal) 5.08 30.70 5.85 β = 1.5 5.12 30.61 5.80 β = 1.9 (Near MMD-like) 5.18 30.48 5.75 Time Weighting w(t) w(t) = 1.0 (Uniform) 5.25 30.33 5.65 w(t) = σ2 t (Optimal, VSD-like) 5.08 30.70 5.85 w(t) = 1/σt (Score-like) 5.18 30.58 5.77 w(t) = αt/σ2 t (SNR-based) 5.11 30.65 5.82 Batch Size K (per GPU) K = 4 5.28 30.25 5.60 K = 8 5.16 30.50 5.76 K = 16 (Optimal) 5.08 30.70 5.85 K = 32 5.05 30.75 5.82 |