Pith. sign in

REVIEW 2 major objections 5 minor 80 references

FourTune makes end-to-end 4-bit post-training of large diffusion models match full-precision quality while cutting memory and raising throughput.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

FourTune matches full-precision LoRA quality on diffusion post-training via native W4A4G4 with a frozen SVD stabilizer, block-wise quant, and fused kernels, cutting memory 2.25× and speeding training 2.27× on FLUX.1-dev.

T0 review reviewed 2026-07-11 challenge →

load-bearing objection Solid systems result: first stable end-to-end W4A4G4 post-training for large diffusion models that matches BF16 LoRA quality while beating the QLoRA memory-speed trade-off on Blackwell; main soft spot is under-specified stabilizer rank, not a broken claim. the 2 major comments →

arxiv 2607.05711 v1 pith:QKRTTKWU submitted 2026-07-07 cs.LG cs.CV

FourTune: Towards Fully 4-Bit Efficient Post-Training for Diffusion Models

Bowen Xue , Zihan Min , Xingyang Li , Zhekai Zhang , Haocheng Xi , Lvmin Zhang , Maneesh Agrawala , Jun-Yan Zhu
show 3 more authors
Song Han Yujun Lin Muyang Li
This is my paper
classification cs.LG cs.CV
keywords diffusion modelspost-training4-bit quantizationW4A4G4LoRAparameter-efficient fine-tuningnumerical stabilizerblock-wise quantization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Post-training large diffusion models for customization, preference alignment, and distillation is still limited by memory and speed even when only low-rank adapters are trained. FourTune claims that an end-to-end W4A4G4 pipeline can remove that bottleneck: weights, activations, and gradients of the frozen backbone all run in 4-bit arithmetic. A frozen low-rank full-precision stabilizer, obtained by SVD of the pretrained weights, isolates quantization-sensitive outliers so the residual can stay in native 4-bit without gradient explosion. Block-wise quantization and fused kernels then make the 4-bit backward pass and LoRA/MLP paths bandwidth-efficient. Across the three post-training regimes the method matches BF16 LoRA quality; on the 12B FLUX.1-dev model it reports 2.25 imes lower memory and 2.27 imes higher end-to-end step throughput than BF16 LoRA.

Core claim

A triple-branch hybrid pipeline—frozen 4-bit residual backbone, frozen full-precision low-rank stabilizer, and trainable LoRA—makes native W4A4G4 forward and backward passes numerically stable for large diffusion models, so that post-training quality matches full-precision LoRA while memory falls and training throughput rises substantially.

What carries the argument

The triple-branch hybrid-precision pipeline: after SVD decomposition W ≈ R + L_stab, the residual R is quantized to 4-bit, L_stab is kept frozen in high precision as a numerical stabilizer, and a trainable LoRA branch adapts the model; block-wise 16×16 quantization plus fused kernels then enable direct 4-bit transposed GEMMs in the backward pass.

Load-bearing premise

A frozen low-rank full-precision stabilizer taken from an SVD of the pretrained weights is enough to keep native 4-bit forward and backward passes stable and able to reach full-precision task quality.

What would settle it

Remove the stabilizer branch, retrain under the same W4A4G4 schedule on the reported customization, RL or distillation tasks, and check whether gradient norms explode (as in the paper’s ablation) and final similarity, aesthetic, or FID metrics fall well below the BF16 LoRA baseline.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Post-training of 12B–20B diffusion models becomes feasible on consumer Blackwell GPUs with roughly half the memory of BF16 LoRA.
  • Customization, preference RL, and few-step distillation can all be run under the same W4A4G4 stack without separate precision recipes.
  • Once the stabilizer is pre-computed, the dominant backbone arithmetic stays in 4-bit Tensor Cores for both forward and backward passes.
  • Block-wise quantization removes the need to dequantize and re-quantize weights when forming the transposed product in the backward pass.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same stabilizer-plus-residual pattern may transfer to other PEFT methods beyond LoRA if their adapters remain low-rank and full-precision.
  • If the SVD stabilizer rank is small enough, the approach could also shrink the memory of full-parameter fine-tuning by quantizing only the residual.
  • Hardware that already ships native 4-bit GEMMs (Blackwell, future consumer parts) becomes the practical enabler; older architectures would need a different kernel stack.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 5 minor

Summary. FourTune proposes an end-to-end W4A4G4 post-training framework for large diffusion models. It augments LoRA with a frozen full-precision low-rank stabilizer obtained by SVD of the pretrained weights (Eq. 1–2, §3.1), so that the residual backbone can be quantized to NVFP4 and executed natively in 4-bit for both forward and backward passes. Block-wise (16×16) quantization enables efficient online transposition of 4-bit weights during backpropagation, and fused kernels reduce memory traffic for the LoRA and MLP paths. Experiments on customization (FLUX.1-dev 12B and Qwen-Image 20B), SRPO reinforcement learning, and π-Flow distillation report generation quality matching BF16 LoRA, with claimed 2.25× memory reduction and 2.27× end-to-end training throughput on FLUX.1-dev versus BF16 LoRA. Ablations address precision configurations, stabilizer presence, quantization granularity, and kernel fusion.

Significance. If the quality-matching and efficiency claims hold under fully specified hyperparameters, the work would be a useful systems contribution: it is the first reported native W4A4G4 post-training pipeline for large generative models, and it demonstrates practical speedups on Blackwell hardware while remaining competitive with BF16 LoRA across three distinct post-training regimes. Strengths include multi-task evaluation (customization, RL, distillation), multi-model scale (12B and 20B), explicit efficiency breakdowns (memory and latency), and ablations that isolate the stabilizer, block-wise quantization, and kernel fusion. The approach builds cleanly on SVDQuant for the residual/stabilizer split and on hardware-friendly formats (NVFP4/INT4). The main open question is whether the quality match is robust to the (currently unreported) stabilizer rank, which determines residual dynamic range and thus how much of the numerical burden is carried by the frozen full-precision path.

major comments (2)
  1. [§3.1, Eq. (1)–(2); Fig. 8] §3.1, Eq. (1)–(2) and the stabilizer ablation (Fig. 8): the frozen low-rank stabilizer L_stab = L1 L2 is load-bearing for numerical stability under native 4-bit training; without it, gradient norms explode. The manuscript never states the SVD rank (or energy retained) used for L_stab, nor residual max/scale statistics after decomposition. Without these numbers it is impossible to judge whether residual 4-bit representability is robust or an artifact of a high-rank full-precision path that absorbs most outliers. Please report rank (or rank schedule), residual dynamic-range statistics, and at least a one-dimensional sensitivity study over stabilizer rank on one task (e.g., identity customization or the distillation gradient-norm setting).
  2. [Abstract; Fig. 1; §4.2] Efficiency claims (abstract, Fig. 1, §4.2): the headline 2.25× memory and 2.27× throughput figures are end-to-end on FLUX customization, while DiT-only numbers are stronger (2.93× weight memory, 2.52× step latency). The contribution of the frozen stabilizer’s full-precision storage and GEMMs to the memory and latency budgets is not broken out. Please add a memory/latency decomposition that isolates residual 4-bit weights, stabilizer, LoRA adapters, activations, and optimizer state so that the net gain of the triple-branch design versus pure W4A4G4 (or versus QLoRA) is transparent.
minor comments (5)
  1. [Table 1] Table 1: several metrics for Ours are slightly better or worse than BF16 LoRA without error bars or multi-seed statistics; a short note on variance (or at least that numbers are single-run) would strengthen the “matches quality” claim.
  2. [§3.2] §3.2: the argument that block-wise scales remain valid after online transposition is clear, but a short pseudocode or kernel sketch of the transposed 4-bit GEMM would help readers implement the idea.
  3. [Appendix B] Appendix B (SDXL and INT4/RTX 4090) is valuable for generality; consider moving a one-sentence summary into the main text so the claim is not buried.
  4. [Fig. 3] Figure 3 caption and surrounding text: clarify whether the stabilizer is applied to every linear layer or only selected modules (e.g., attention vs. MLP).
  5. [§3.1] Minor notation: Q(·) is used for both offline residual quantization and dynamic activation quantization; a brief distinction would avoid ambiguity.

Circularity Check

1 steps flagged

No significant circularity: empirical systems paper whose quality/speed claims are measured against external baselines; minor self-citation of SVDQuant decomposition as a reusable building block, not a load-bearing uniqueness claim.

specific steps
  1. self citation load bearing [§3.1, Eq. (1) and surrounding text]
    "Formally, building upon the spectral decomposition from SVDQuant (Li et al., 2025), we decompose the pre-trained weight matrix W∈R m×n into a quantization-friendly residual R and a low-rank outlier component L stab: W≈R+L stab,where L stab =L 1L2."

    The stabilizer construction that makes native W4A4G4 training stable is taken from prior work by overlapping authors (SVDQuant). The citation is not a uniqueness theorem and the paper supplies its own ablation (Fig. 8) showing necessity; still, the load-bearing numerical-stability premise rests partly on that self-citation rather than being derived independently here. Minor and non-central to the empirical quality claims.

full rationale

FourTune is an engineering/systems paper. Its central claims (W4A4G4 quality matching BF16 LoRA on customization/RL/distillation, plus measured 2.25× memory and 2.27× throughput gains) are established by direct experimental comparison to external full-precision and QLoRA baselines using public-style metrics (AntelopeV2, CLIP, DINOv3, HPSv2.1, FID, etc.). There is no first-principles derivation, fitted constant renamed as prediction, or self-definitional loop. The only self-citation of note is the reuse of the SVD residual+outlier decomposition from SVDQuant (Li et al., 2025, overlapping authors) to construct the frozen stabilizer L_stab in Eq. 1–2. That citation supplies a known technique for isolating outliers; the paper then shows empirically (Fig. 8 ablation) that the stabilizer prevents gradient explosion under native 4-bit training and that the overall pipeline matches full-precision task quality. This is ordinary method reuse, not a circular reduction of the result to its inputs, nor an imported uniqueness theorem that forbids alternatives. Stabilizer rank is under-specified (a correctness/reproducibility concern), but that does not create circularity. Score 1 reflects only the minor overlapping-author citation of a prior method; the derivation chain itself is non-circular and self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central claim rests on standard PEFT/quantization practice plus a few design choices (stabilizer from SVD, 16×16 blocks, LoRA rank, NVFP4) and the empirical premise that 4-bit GEMM plus a frozen outlier branch is stable enough for diffusion post-training. No new physical entities; free parameters are engineering knobs; axioms are domain assumptions about hardware and quantization error, not ad-hoc mathematical postulates invented to force a theorem.

free parameters (4)
  • LoRA rank r
    Fixed at r=64 for main customization/RL/distillation runs; controls adapter capacity and is a hand-chosen PEFT hyperparameter the quality claims depend on matching BF16 LoRA at the same rank.
  • Block-wise quantization granularity
    Chosen as 16×16 blocks to enable efficient transposed 4-bit GEMM in the backward pass; coarser than 1×16 group-wise and is a design knob validated only by LPIPS/PSNR on MJHQ-30K reconstruction, not derived.
  • Stabilizer rank / SVD truncation
    Low-rank L_stab = L1 L2 from SVD of pretrained W; the retained rank is a free design choice that trades residual quantizability against stabilizer cost and is not uniquely determined by theory in the paper.
  • Quantization format and scales (NVFP4 / INT4)
    Hardware-facing 4-bit format and per-block scales are implementation choices; main speed claims assume Blackwell NVFP4 tensor cores.
axioms (4)
  • domain assumption Low-rank adapters (LoRA) can match full fine-tuning quality for diffusion post-training when the backbone is frozen.
    Standard PEFT premise used throughout; quality comparisons are always vs BF16 LoRA, not full FT except in the RL table.
  • domain assumption SVD can separate quantization-sensitive outliers into a small full-precision factor so the residual is 4-bit friendly (from SVDQuant).
    Eq. (1) and §3.1 import this decomposition; training stability claims rest on it remaining valid under gradient flow.
  • domain assumption NVIDIA Blackwell (and related) 4-bit tensor cores correctly implement the intended W4A4G4 GEMMs used for timing.
    Efficiency claims in §4.2 are measured on RTX Pro 6000 / 5090; arithmetic correctness of the hardware path is assumed.
  • ad hoc to paper Block-local scales remain valid after online transposition of 4-bit weight blocks.
    §3.2 design justification for 16×16 blocks; required for the claimed dequant-free 4-bit backward path.
invented entities (1)
  • Frozen numerical stabilizer branch in a triple-branch W4A4G4 training pipeline no independent evidence
    purpose: Isolate outliers so the quantized residual can run native 4-bit forward and backward without gradient explosion while LoRA adapts the task.
    New as a training-time frozen pathway (SVDQuant used a similar split for inference). Independent evidence is only the paper’s own gradient-norm ablation and task metrics, not an external falsifiable prediction.

reviewed 2026-07-11 · how reviews work

0 comments
Cite this review

Pith. "Pith review of FourTune: Towards Fully 4-Bit Efficient Post-Training for Diffusion Models." pith.science (2026). https://pith.science/paper/QKRTTKWU

@misc{pith2026260705711,
  author       = {Pith},
  title        = {Pith review of: FourTune: Towards Fully 4-Bit Efficient Post-Training for Diffusion Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QKRTTKWU}},
  note         = {Machine review of arXiv:2607.05711}
}
Share X Bluesky LinkedIn Reddit HN
abstract

Diffusion models have become a dominant paradigm for high-quality generative modeling, while post-training is essential for adapting them to diverse downstream applications. However, post-training of large diffusion models is still challenging due to the prohibitive memory footprints and slow training speed, which existing parameter-efficient fine-tuning methods only partially address. To overcome these limitations, we propose FourTune, an efficient post-training framework for diffusion models based on an end-to-end W4A4G4 paradigm. FourTune introduces a triple-branch hybrid pipeline that augments the standard LoRA architecture with a frozen numerical stabilizer to isolate quantization-sensitive outliers, enabling stable training under native 4-bit computation. In addition, FourTune employs hardware-efficient block-wise quantization and customized fused kernels to support efficient quantized backpropagation and reduce memory bandwidth overhead. Across customization, reinforcement learning, and distillation tasks, FourTune matches the quality of full-precision fine-tuning. On FLUX.1-dev (12B), FourTune reduces memory overhead by 2.25$\times$ and increases end-to-end training throughput by 2.27$\times$ compared to BF16 LoRA.

Figures

Figures reproduced from arXiv: 2607.05711 by Bowen Xue, Haocheng Xi, Jun-Yan Zhu, Lvmin Zhang, Maneesh Agrawala, Muyang Li, Song Han, Xingyang Li, Yujun Lin, Zhekai Zhang, Zihan Min.

Figure 1
Figure 1. Figure 1: Qualitative and quantitative comparison of FourTune against baselines. Top: Visual comparisons across three diverse post-training tasks: Customization, Reinforcement Learning, and Distillation. Despite extremely low-bit quantization (W4A4G4), FourTune produces high-fidelity images visually indistinguishable from the full-precision BF16 LoRA baseline. Bottom: Efficiency benchmarks performed on FLUX.1-dev (1… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of forward and backward pipelines across different PEFT methods. (a) LoRA follows the standard high-precision training pipeline, where weights (W), activations (A), and gradients (G) are all stored and computed in high precision. (b) QLoRA reduces the memory footprint by storing weights in 4-bit precision, but still requires on-the-fly dequantization to 16-bit for computation (W4A16G16). (c) Our… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the FourTune framework. (a) Weight Quantization: High-precision weights are decomposed via SVD into a quantization-friendly 4-bit residual and a high-precision stabilizer. (b) Triple-Branch Pipeline: The architecture consists of a frozen 4-bit backbone for arithmetic efficiency, a frozen stabilizer to ensure numerical precision, and a trainable LoRA branch for task adaptation. 3. Method We prop… view at source ↗
Figure 4
Figure 4. Figure 4: Validation curves for customization tasks, showing per￾formance on par with full-precision training and with substantial acceleration. Reinforcement Learning Setup. We evaluate FourTune in a reinforcement learning setting using SRPO (Fei et al., 5 [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison on customization. FourTune matches the generation quality of full-precision LoRA. style transfer, and general subject reconstruction, validating its ability to capture fine-grained visual details. FLUX.1-dev Full Finetune BF16 LoRA Ours RL Compariso NF4 QLoRA n BF16 Finetune [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison on RL. FourTune even matches the performance of full-precision, full-parameter finetuning. For Reinforcement Learning ( [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: Effect of the Stabilizer on numerical stability. Com￾parison of gradient norms with and without the proposed Stabilizer. The results show that our method prevents the gradient explosion observed in the baseline setting (red line), ensuring stable conver￾gence. Effectiveness of the Stabilizer. During native 4-bit train￾ing, the accumulation of quantization errors occasionally triggers gradient explosion, re… view at source ↗
Figure 9
Figure 9. Figure 9: Breakdown of efficiency gains from different kernel fusion strategies. The combined optimization results in a 2.52× speedup relative to 16-bit LoRA fine-tuning. 5. Conclusion We have presented FourTune, a fully 4-bit post-training framework that enables end-to-end W4A4G4 optimization for large diffusion models by using a frozen stabilizer branch, block-wise quantization for efficient backward com￾putation,… view at source ↗
Figure 10
Figure 10. Figure 10: shows the complete dataset used on customization tasks. The dataset consists of three categories: Human Identity, which focuses on preserving personal identity while allowing variations in appearance or attributes. Artistic Style, which targets adapting artistic or visual styles while keeping the underlying content unchanged. General Subject, which emphasizes generating a specific subject consistently acr… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

80 extracted references · 80 canonical work pages

  1. [1]

    High-Resolution Image Synthesis with Latent Diffusion Models , booktitle = CVPR, year =

    Robin Rombach and Andreas Blattmann and Dominik Lorenz and Patrick Esser and Bj. High-Resolution Image Synthesis with Latent Diffusion Models , booktitle = CVPR, year =

    work page

  2. [2]

    Dustin Podell and Zion English and Kyle Lacey and Andreas Blattmann and Tim Dockhorn and Jonas M

    work page

  3. [3]

    Yuwei Guo and Ceyuan Yang and Anyi Rao and Zhengyang Liang and Yaohui Wang and Yu Qiao and Maneesh Agrawala and Dahua Lin and Bo Dai , title =

    work page

  4. [4]

    ArXiv preprint , year =

    Andreas Blattmann and Tim Dockhorn and Sumith Kulal and Daniel Mendelevitch and Maciej Kilian and Dominik Lorenz and Yam Levi and Zion English and Vikram Voleti and Adam Letts and Varun Jampani and Robin Rombach , title =. ArXiv preprint , year =

    work page

  5. [5]

    William Peebles and Saining Xie , title =

    work page

  6. [6]

    Kwok and Ping Luo and Huchuan Lu and Zhenguo Li , title =

    Junsong Chen and Jincheng Yu and Chongjian Ge and Lewei Yao and Enze Xie and Zhongdao Wang and James T. Kwok and Ping Luo and Huchuan Lu and Zhenguo Li , title =

    work page

  7. [7]

    Qwen-Image Technical Report , journal =

    work page

  8. [8]

    2025 , howpublished=

    work page 2025

  9. [9]

    Enze Xie and Junsong Chen and Junyu Chen and Han Cai and Haotian Tang and Yujun Lin and Zhekai Zhang and Muyang Li and Ligeng Zhu and Yao Lu and Song Han , title =

    work page

  10. [10]

    Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer , journal =

    work page

  11. [11]

    Wan: Open and Advanced Large-Scale Video Generative Models , journal =

    work page

  12. [12]

    Zhuoyi Yang and Jiayan Teng and Wendi Zheng and Ming Ding and Shiyu Huang and Jiazheng Xu and Yuanming Yang and Wenyi Hong and Xiaohan Zhang and Guanyu Feng and Da Yin and Yuxuan Zhang and Weihan Wang and Yean Cheng and Bin Xu and Xiaotao Gu and Yuxiao Dong and Jie Tang , title =

    work page

  13. [13]

    2024 , journal =

    HunyuanVideo:. 2024 , journal =

    work page 2024

  14. [14]

    HunyuanVideo 1.5 Technical Report , journal =

    work page

  15. [15]

    Multi-Concept Customization of Text-to-Image Diffusion , booktitle = CVPR, year =

    Nupur Kumari and Bingliang Zhang and Richard Zhang and Eli Shechtman and Jun. Multi-Concept Customization of Text-to-Image Diffusion , booktitle = CVPR, year =

    work page

  16. [16]

    Nataniel Ruiz and Yuanzhen Li and Varun Jampani and Yael Pritch and Michael Rubinstein and Kfir Aberman , title =

    work page

  17. [17]

    An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion , booktitle = ICLR, year =

    Rinon Gal and Yuval Alaluf and Yuval Atzmon and Or Patashnik and Amit Haim Bermano and Gal Chechik and Daniel Cohen. An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion , booktitle = ICLR, year =

    work page

  18. [18]

    Fleet , title =

    Kevin Clark and Paul Vicol and Kevin Swersky and David J. Fleet , title =

    work page

  19. [19]

    Aligning Text-to-Image Diffusion Models with Reward Backpropagation , journal =

    Mihir Prabhudesai and Anirudh Goyal and Deepak Pathak and Katerina Fragkiadaki , year =. Aligning Text-to-Image Diffusion Models with Reward Backpropagation , journal =

    work page

  20. [20]

    ArXiv preprint , year =

    Mihir Prabhudesai and Russell Mendonca and Zheyang Qin and Katerina Fragkiadaki and Deepak Pathak , title =. ArXiv preprint , year =

    work page

  21. [21]

    Jiazheng Xu and Xiao Liu and Yuchen Wu and Yuxuan Tong and Qinkai Li and Ming Ding and Jie Tang and Yuxiao Dong , title =

    work page

  22. [22]

    ArXiv preprint , year =

    Xiangwei Shen and Zhimin Li and Zhantao Yang and Shiyi Zhang and Yingfang Zhang and Donghao Li and Chunyu Wang and Qinglin Lu and Yansong Tang , title =. ArXiv preprint , year =

    work page

  23. [23]

    2025 , journal =

    SRPO: Self-Referential Policy Optimization for Vision-Language-Action Models , author=. 2025 , journal =

    work page 2025

  24. [24]

    Kevin Black and Michael Janner and Yilun Du and Ilya Kostrikov and Sergey Levine , title =

    work page

  25. [25]

    Optimizing

    Ying Fan and Kangwook Lee , editor =. Optimizing

    work page

  26. [26]

    Ying Fan and Olivia Watkins and Yuqing Du and Hao Liu and Moonkyung Ryu and Craig Boutilier and Pieter Abbeel and Mohammad Ghavamzadeh and Kangwook Lee and Kimin Lee , title =

    work page

  27. [27]

    2025 , journal =

    Jie Liu and Gongye Liu and Jiajun Liang and Yangguang Li and Jiaheng Liu and Xintao Wang and Pengfei Wan and Di Zhang and Wanli Ouyang , title =. 2025 , journal =

    work page 2025

  28. [28]

    ArXiv preprint , year =

    Yibin Wang and Zhimin Li and Yuhang Zang and Yujie Zhou and Jiazi Bu and Chunyu Wang and Qinglin Lu and Cheng Jin and Jiaqi Wang , title =. ArXiv preprint , year =

    work page

  29. [29]

    ArXiv preprint , year =

    Zeyue Xue and Jie Wu and Yu Gao and Fangyuan Kong and Lingting Zhu and Mengzhao Chen and Zhiheng Liu and Wei Liu and Qiushan Guo and Weilin Huang and Ping Luo , title =. ArXiv preprint , year =

    work page

  30. [30]

    ArXiv preprint , year =

    Eric Luhman and Troy Luhman , title =. ArXiv preprint , year =

    work page

  31. [31]

    Tim Salimans and Thomas Mensink and Jonathan Heek and Emiel Hoogeboom , title =

    work page

  32. [32]

    Xingchao Liu and Chengyue Gong and Qiang Liu , title =

    work page

  33. [33]

    Xingchao Liu and Xiwen Zhang and Jianzhu Ma and Jian Peng and Qiang Liu , title =

    work page

  34. [34]

    Kevin Frans and Danijar Hafner and Sergey Levine and Pieter Abbeel , title =

    work page

  35. [35]

    Guibas and Gordon Wetzstein and Sai Bi , title =

    Hansheng Chen and Kai Zhang and Hao Tan and Leonidas J. Guibas and Gordon Wetzstein and Sai Bi , title =. ArXiv preprint , year =

    work page

  36. [36]

    Song, Yang and Dhariwal, Prafulla and Chen, Mark and Sutskever, Ilya , title =

    work page

  37. [37]

    Susskind , title =

    Jiatao Gu and Chen Wang and Shuangfei Zhai and Yizhe Zhang and Lingjie Liu and Joshua M. Susskind , title =

    work page

  38. [38]

    Consistency Trajectory Models: Learning Probability Flow

    Dongjun Kim and Chieh. Consistency Trajectory Models: Learning Probability Flow

    work page

  39. [39]

    Yang Song and Prafulla Dhariwal , title =

    work page

  40. [40]

    Zico Kolter , title =

    Zhengyang Geng and Ashwini Pokle and Weijian Luo and Justin Lin and J. Zico Kolter , title =

    work page

  41. [41]

    Flow map matching with stochastic interpolants:

    Nicholas Matthew Boffi and Michael Samuel Albergo and Eric Vanden. Flow map matching with stochastic interpolants:. Trans. Mach. Learn. Res. , year =

    work page

  42. [42]

    One-Step Diffusion with Distribution Matching Distillation , booktitle = CVPR, year =

    Tianwei Yin and Micha. One-Step Diffusion with Distribution Matching Distillation , booktitle = CVPR, year =

    work page

  43. [43]

    Improved Distribution Matching Distillation for Fast Image Synthesis , booktitle = NeurIPS, year =

    Tianwei Yin and Micha. Improved Distribution Matching Distillation for Fast Image Synthesis , booktitle = NeurIPS, year =

    work page

  44. [44]

    Axel Sauer and Dominik Lorenz and Andreas Blattmann and Robin Rombach , title =

    work page

  45. [45]

    Mingyuan Zhou and Huangjie Zheng and Zhendong Wang and Mingzhang Yin and Hai Huang , title =

    work page

  46. [46]

    Zico Kolter and Guo

    Weijian Luo and Zemin Huang and Zhengyang Geng and J. Zico Kolter and Guo. One-Step Diffusion Distillation through Score Implicit Matching , booktitle = NeurIPS, year =

    work page

  47. [47]

    Linqi Zhou and Stefano Ermon and Jiaming Song , title =

    work page

  48. [48]

    Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen

    Edward J. Hu and Yelong Shen and Phillip Wallis and Zeyuan Allen. LoRA: Low-Rank Adaptation of Large Language Models , booktitle = ICLR, year =

    work page

  49. [49]

    ArXiv preprint , year =

    Vera Soboleva and Aibek Alanov and Andrey Kuznetsov and Konstantin Sobolev , title =. ArXiv preprint , year =

    work page

  50. [50]

    Qingru Zhang and Minshuo Chen and Alexander Bukharin and Pengcheng He and Yu Cheng and Weizhu Chen and Tuo Zhao , title =

    work page

  51. [51]

    Soufiane Hayou and Nikhil Ghosh and Bin Yu , title =

    work page

  52. [52]

    DoRA: Weight-Decomposed Low-Rank Adaptation , booktitle = ICML, year =

    Shih. DoRA: Weight-Decomposed Low-Rank Adaptation , booktitle = ICML, year =

    work page

  53. [53]

    2025 , url =

    Schulman, John and. 2025 , url =

    work page 2025

  54. [54]

    Xiuyu Li and Yijiang Liu and Long Lian and Huanrui Yang and Zhen Dong and Daniel Kang and Shanghang Zhang and Kurt Keutzer , title =

    work page

  55. [55]

    Yuzhang Shang and Zhihang Yuan and Bin Xie and Bingzhe Wu and Yan Yan , title =

    work page

  56. [56]

    SVDQuant: Absorbing Outliers by Low-Rank Component for 4-Bit Diffusion Models , booktitle = ICLR, year =

    Muyang Li and Yujun Lin and Zhekai Zhang and Tianle Cai and Xiuyu Li and Junxian Guo and Enze Xie and Chenlin Meng and Jun. SVDQuant: Absorbing Outliers by Low-Rank Component for 4-Bit Diffusion Models , booktitle = ICLR, year =

    work page

  57. [57]

    ArXiv preprint , year =

    NVIDIA , title =. ArXiv preprint , year =

    work page

  58. [58]

    ArXiv preprint , year =

    Wei Huang and Yi Ge and Shuai Yang and Yicheng Xiao and Huizi Mao and Yujun Lin and Hanrong Ye and Sifei Liu and Ka Chun Cheung and Hongxu Yin and Yao Lu and Xiaojuan Qi and Song Han and Yukang Chen , title =. ArXiv preprint , year =

    work page

  59. [59]

    Tim Dettmers and Artidoro Pagnoni and Ari Holtzman and Luke Zettlemoyer , title =

    work page

  60. [60]

    Yixiao Li and Yifan Yu and Chen Liang and Nikos Karampatziakis and Pengcheng He and Weizhu Chen and Tuo Zhao , title =

    work page

  61. [61]

    Yuhui Xu and Lingxi Xie and Xiaotao Gu and Xin Chen and Heng Chang and Hengheng Zhang and Zhengsu Chen and Xiaopeng Zhang and Qi Tian , title =

    work page

  62. [62]

    2022 , howpublished =

    InsightFace Python Library (Model Zoo: antelopev2) , author =. 2022 , howpublished =

    work page 2022

  63. [63]

    ArcFace: Additive Angular Margin Loss for Deep Face Recognition , author =

    work page

  64. [64]

    Alec Radford and Jong Wook Kim and Chris Hallacy and Aditya Ramesh and Gabriel Goh and Sandhini Agarwal and Girish Sastry and Amanda Askell and Pamela Mishkin and Jack Clark and Gretchen Krueger and Ilya Sutskever , title =

    work page

  65. [65]

    DINOv3 , journal =

    Oriane Sim. DINOv3 , journal =

    work page

  66. [66]

    ArXiv preprint , year =

    Xiaoshi Wu and Yiming Hao and Keqiang Sun and Yixiong Chen and Feng Zhu and Rui Zhao and Hongsheng Li , title =. ArXiv preprint , year =

    work page

  67. [67]

    2024 , howpublished =

    discus0434 , title =. 2024 , howpublished =

    work page 2024

  68. [68]

    Yuval Kirstain and Adam Polyak and Uriel Singer and Shahbuland Matiana and Joe Penna and Omer Levy , title =

    work page

  69. [69]

    ImageReward: Learning and Evaluating Human Preferences for Text-to-Image Generation , author =

    work page

  70. [70]

    Zhiqiu Lin and Deepak Pathak and Baiqi Li and Jiayao Li and Xide Xia and Graham Neubig and Pengchuan Zhang and Deva Ramanan , title =

    work page

  71. [71]

    Howard and Hartwig Adam and Dmitry Kalenichenko , title =

    Benoit Jacob and Skirmantas Kligys and Bo Chen and Menglong Zhu and Matthew Tang and Andrew G. Howard and Hartwig Adam and Dmitry Kalenichenko , title =

    work page

  72. [72]

    2024 , url=

    NVIDIA Blackwell Architecture Technical Brief , author=. 2024 , url=

    work page 2024

  73. [73]

    Microscaling Data Formats for Deep Learning , journal =

    Bita Darvish Rouhani and Ritchie Zhao and Ankit More and Mathew Hall and Alireza Khodamoradi and Summer Deng and Dhruv Choudhary and Marius Cornea and Eric Dellinger and Kristof Denolf and Dusan Stosic and Venmugil Elango and Maximilian Golub and Alexander Heinecke and Phil James. Microscaling Data Formats for Deep Learning , journal =

    work page

  74. [74]

    Navigating Text-To-Image Customization: From Ly

    SHIH-YING YEH and Yu-Guan Hsieh and Zhidong Gao and Bernard B W Yang and Giyeong Oh and Yanmin Gong , booktitle=ICLR, year=. Navigating Text-To-Image Customization: From Ly

    work page

  75. [75]

    Chaofeng Chen and Jiadi Mo , year=

    work page

  76. [76]

    Junnan Li and Dongxu Li and Caiming Xiong and Steven C. H. Hoi , editor =

    work page

  77. [77]

    International Joint Conference on Natural Language Processing , year =

    Armen Aghajanyan and Sonal Gupta and Luke Zettlemoyer , title =. International Joint Conference on Natural Language Processing , year =

    work page

  78. [78]

    Neil Houlsby and Andrei Giurgiu and Stanislaw Jastrzebski and Bruna Morrone and Quentin de Laroussilhe and Andrea Gesmundo and Mona Attariyan and Sylvain Gelly , title =

    work page

  79. [79]

    ArXiv preprint , year =

    Chengsong Huang and Qian Liu and Bill Yuchen Lin and Tianyu Pang and Chao Du and Min Lin , title =. ArXiv preprint , year =

    work page

  80. [80]

    Plataniotis and Amir Yazdanbakhsh and Kurt Keutzer and Song Han and Zhijian Liu , title =

    Samir Khaki and Xiuyu Li and Junxian Guo and Ligeng Zhu and Konstantinos N. Plataniotis and Amir Yazdanbakhsh and Kurt Keutzer and Song Han and Zhijian Liu , title =

    work page

This paper was first reviewed by grok-4.5 on July 11, 2026.