Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

A compact 4B stack—one-step VAE, native-resolution transformer, fused kernels—matches 20B+ systems at a fraction of the cost.

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 ↗

T0 review · deepseek-v4-flash

2026-08-01 13:32 UTC pith:XZZK36CD

load-bearing objection Strong systems paper; the tokenizer is the real contribution, the 'surpasses larger models' framing isn't backed by the reported numbers. the 4 major comments →

arxiv 2607.19064 v2 pith:XZZK36CD submitted 2026-07-21 cs.CV cs.AIcs.LGcs.MMeess.IV

Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing

Xinjie Zhang , Peng Zhang , Shicheng Zheng , Jinghao Guo , Zhaoyang Jia , Yifei Shen , Xun Guo , Yuxuan Luo
show 16 more authors
Jiahao Li Wenxuan Xie Fanyi Pu Xiaoyi Zhang Kaichen Zhang Zongyu Guo Tianci Bi Dongnan Gui Zhening Liu Zimo Wen Zihan Zheng Senqiao Yang Xiao Li Jinglu Wang Bin Li Yan Lu
This is my paper
classification cs.CV cs.AIcs.LGcs.MMeess.IV
keywords image generationimage editinglatent tokenizerdiffusion transformerrectified flowfew-step distillationnative resolutionefficient inference
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 paper argues that strong image generation and editing do not require tens-of-billions-parameter backbones. It introduces a 4B-parameter stack whose key move is replacing the heavy VAE tokenizer with a one-step, fully convolutional encoder and decoder, regularized toward the latent distribution of a strong public VAE so the latent space stays generation-ready. On top of this, a native-resolution diffusion transformer is trained with packed variable-length sequences, removing resolution buckets and speeding end-to-end training by roughly 2.5x through fused kernels. The resulting family, including aligned and 4-step Turbo variants, matches or surpasses much larger open-source systems on standard benchmarks while running 1024x1024 generation in 0.59s and editing in 1.02s on a single A100 GPU with about 18-20GB peak memory. If right, this makes high-quality visual generation practical to study, fine-tune, and deploy under realistic compute budgets.

Core claim

The central claim is that a carefully co-designed 4B-scale stack—a lightweight one-step diffusion-style VAE, a native-resolution multimodal diffusion transformer, and fused-kernel training infrastructure—can deliver image generation and instruction-based editing competitive with 20B+ open-source systems. The load-bearing quantitative claims are that Mage-VAE attains reconstruction fidelity on par with the strongest public VAE while requiring roughly 12x and 22x fewer encoding and decoding MACs per pixel, and that the Turbo variants run 4-step generation at 0.59s and 4-step editing at 1.02s at 1024x1024 on a single A100 GPU with about 18-20GB peak memory. The paper also reports that online po

What carries the argument

The central object is Mage-VAE, a latent tokenizer built as the architectural dual of a one-step pixel-diffusion decoder: both encoder and decoder are fully convolutional one-step diffusion models with no global attention, so their cost stays roughly linear in resolution. The trick that keeps the lightweight latent space usable for generation is anchor-latent KL regularization: the encoder's posterior is pulled toward the patchified latent distribution of a strong public VAE (16x spatial reduction, 128 channels), letting Mage-VAE be swapped into existing generators without retraining them. Around this, native-resolution packing packs variable-length image and text sequences into one batch us

Load-bearing premise

The claim that the 4B model matches or surpasses much larger systems rests on automated judge scores (an LLM-based visual judge, an OCR model, and a reasoning reward model) ranking model quality faithfully and on single-run benchmark numbers being stable; if those judges are noisy or benchmark selection flips the small deltas, the efficiency-quality trade-off survives but the specific 'surpasses' framing weakens.

What would settle it

Run GenEval, GEdit, ImgEdit, and TextEdit with the released checkpoints multiple times, varying judge seeds and temperature or using two independent judge models, and compute confidence intervals; if the 4B model's scores fall below the 20B systems by more than the interval in a majority of runs, the 'competitive or surpasses' central claim is refuted. A cheaper check: measure Mage-VAE encoding and decoding MACs and latency on a 4096x4096 image—if it is not roughly an order of magnitude faster than the public VAE it anchors to, the efficiency claim collapses.

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

If this is right

  • A 4B open generator can match or beat 20B+ systems on GenEval, GEdit, and several text-rendering benchmarks, so parameter count is not the only axis of generation quality.
  • The VAE bottleneck in high-resolution few-step generation can be removed: one-step tokenization keeps 4-step 1024x1024 generation under a second and makes repeated editing much cheaper.
  • Native-resolution packing lets a single checkpoint output any aspect ratio from 512 to 2048 without resolution buckets, and it also speeds classifier-free guidance inference by about 1.1x.
  • Training throughput improves by about 2.5x and peak memory drops when tokenizer, text encoder, and backbone operator chains are fused, making 4B-scale native-resolution training practical on a single 8-GPU node.
  • Mixing generation data into editing training and adding adversarial perceptual guidance mainly benefit the 4-step Turbo models, not uniformly every editing benchmark.

Where Pith is reading between the lines

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

  • A direct extension the paper leaves implicit: the anchor-latent recipe could distill any expensive VAE into a one-step codec without breaking downstream generator compatibility, potentially helping video or multi-image editing where tokenization is repeated many times.
  • The 'surpasses 20B' framing is the fragile part of the claim; the robust takeaway is quality on par at a fraction of latency and memory. Re-running the benchmarks with multiple judge seeds and reporting confidence intervals would settle how much of the reported advantage is benchmark noise.
  • Because several editing benchmark deltas are small and mixed, a deployment-minded reader should weight the efficiency and memory gains more heavily than the point estimates when choosing a model.
  • The paper's cross-tokenizer swap results suggest a cheap testable prediction: plugging Mage-VAE into an existing FLUX-style generator should preserve its benchmark scores while cutting tokenization cost by an order of magnitude.

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

4 major / 5 minor

Summary. The paper presents Mage-Flow, a 4B-parameter text-to-image generation and instruction-based editing stack. The stack comprises Mage-VAE, a lightweight one-step diffusion encoder/decoder whose latent space is anchored to FLUX.2-VAE via a KL regularizer; a native-resolution MMDiT backbone trained with rectified flow; and stack-level CUDA kernel fusion. The authors instantiate Base, RL-aligned, and Turbo (4-step) variants for both generation and editing. The central claims are: (i) Mage-VAE matches FLUX.2-VAE reconstruction fidelity while reducing encoding/decoding MACs by ~12x/22x; (ii) the 4B stack achieves competitive or superior quality to much larger open-source systems; and (iii) Turbo variants run in 0.59s per 1024^2 image on a single A100. Evidence includes VAE reconstruction/latency tables, cross-tokenizer swaps, benchmark tables covering prompt following, text rendering, and editing, and ablations for adversarial guidance and generation-data mixing.

Significance. If the results hold, this is a practically important demonstration that tokenizer/backbone/system co-design at 4B scale can approach the quality of 20B-80B open generators while offering large efficiency gains. The efficiency evidence is strong and parameter-free: Table 1 and Table 17 give detailed MACs and latencies, and the cross-tokenizer experiment in Table 2 is a sensible compatibility check. The paper releases code, model weights, and detailed training recipes, which would be valuable to the community. However, the quality side of the trade-off is not statistically grounded: the headline 'matches or surpasses much larger open-source systems' is contradicted in places by the paper's own tables and rests on single-run LLM-judge scores with no error bars. The central efficiency thesis is defensible, but the comparative quality claims need substantial empirical tightening.

major comments (4)
  1. [Abstract and Section 6.2.1, Tables 8-9] The claim that Mage-Flow 'matches or surpasses much larger open-source systems' is not supported by the reported numbers. On DPG-Bench, Mage-Flow scores 86.49 vs Qwen-Image 88.32; on OneIG-EN, 0.536 vs 0.539; on LongText-CN, 0.823 vs 0.946. The paper should either restrict the claim to the benchmarks where it genuinely leads (e.g., GenEval, CVTG-2K, GEdit-CN) or rephrase to 'competitive on most benchmarks, with specific strengths and weaknesses.' As written, the central contribution is overstated.
  2. [Section 6.2, Tables 8, 13-16] All generative quality comparisons are single-run point estimates produced by automated judges (GPT-4.1, GPT-4o, OCR) with no error bars, no repeated evaluations, and no significance testing. Several load-bearing deltas are within plausible judge noise: GEdit-EN 8.271 (Mage-Flow-Edit-Turbo) vs 8.276 (JoyAI-Image-Edit); GenEval 0.90 vs 0.89 (Ovis-U1); OneIG-EN 0.536 vs 0.539 (Qwen-Image). The paper should report variance across repeated judge calls (e.g., different temperatures or bootstrapped samples) and clearly state a threshold below which differences are not meaningful, or explicitly label such differences as ties. Without this, the 'competitive/superior' framing cannot be evaluated.
  3. [Section 5.2 and Appendix D, Tables 11 and 16] Text rendering is optimized with PaddleOCR-VL-1.5 as the RL reward and then evaluated with OCR-based metrics on CVTG-2K and LongText. If the same or a closely related recognizer is used for both training reward and evaluation, the reported text-rendering gains may reflect reward overfitting to the recognizer rather than human-legible quality. The authors should evaluate with an independent OCR system, add a human study on a subset, or at minimum disclose this overlap and discuss its potential effect.
  4. [Section 3.1.2 and Table 2] The cross-tokenizer compatibility result is partly by construction: Mage-VAE is trained to predict patchified FLUX.2-VAE latents and regularized with a KL term toward the FLUX.2-VAE latent distribution. Table 2 therefore demonstrates internal consistency of the anchoring design, not an independent validation that the latent space is 'generation-ready.' This should be described as a sanity check rather than evidence that the latent space is intrinsically superior. The reconstruction and efficiency claims rest on Table 1 and Table 17 and are not affected by this concern.
minor comments (5)
  1. [Table 11, rows for FLUX.2-Klein-Base-4B/9B] The reported numbers for FLUX.2-Klein-Base-4B and FLUX.2-Klein-Base-9B are identical across all columns, which is likely a copy/paste error. Please verify.
  2. [Section 6.1] The evaluation protocol states that results are computed 'using the official evaluation protocols,' but it does not specify which OCR model is used for CVTG-2K/LongText in evaluation. This matters given the reward-overfitting concern above; please state the exact evaluator.
  3. [Section 3.1.1] The phrase 'one-step diffusion-style encoding and decoding' is used before the training stages are introduced; a forward pointer to Section 3.1.3 would improve readability.
  4. [Appendix E, Table 19] The row label 'FLUX.2-klein-Base (Zero-shot) 9B' should probably be 'FLUX.2-Klein-Base-9B' for consistency with the rest of the paper.
  5. [References] Reference [33] is cited as 'FLUX.1 Kontext' while the main text also cites FLUX.2; please ensure the version names in the bibliography match the models actually evaluated.

Circularity Check

1 steps flagged

Mage-VAE's FLUX.2 compatibility is a restatement of its anchor-latent training target; core efficiency/reconstruction and benchmark claims are externally validated.

specific steps
  1. self definitional [Section 3.1.4, 'Generation and editing' (Table 2; also Sec. 3.1.2/Appendix A Stage III)]
    "Since Mage-VAE is distilled from and regularized toward the latent distribution of the high-capacity but computationally expensive FLUX.2-VAE, the two tokenizers should remain compatible when used by downstream generators. Table 2 reports a cross-tokenizer ablation that swaps Mage-VAE and FLUX.2-VAE while keeping the generation or editing backbone fixed."

    The encoder is trained to output patchified FLUX.2 anchor latents (Stage I target z_a = P(E_FLUX(x))) and Stage III applies KL regularization D_KL(q_phi(z|x) || q_a(z|x)) toward the FLUX.2 latent distribution. Thus 'compatibility' with FLUX.2-style generators is the training target itself, not an independent prediction; Table 2 is a sanity check of that objective. The paper even derives the expectation from the design ('should remain compatible') before presenting the numbers. Full-system benchmark results still give independent validation of the latent space, so the circularity is limited to the cross-generator compatibility sub-claim.

full rationale

Most load-bearing claims are self-contained and externally measured: Mage-VAE reconstruction is compared with public VAEs on CLIC 2020 and FFHQ (Table 1); cost is given as kMACs/pixel and A100 latency/memory (Table 1, Fig. 7); generation and editing quality use standard benchmarks with external judges (Tables 8-16); training speedups are measured (Table 4); distillation ablations are provided (Tables 6-7). The one step that reduces to its own input is the claimed cross-generator compatibility of Mage-VAE with FLUX.2. Because the latent space is defined by matching FLUX.2's patchified latents and anchor distribution, interchanging the tokenizers in a FLUX.2-style generator is the success criterion of the anchor-latent objective, and Section 3.1.4 explicitly says the compatibility should follow from that design before presenting Table 2 as evidence. This is a definitional restatement rather than an external confirmation, but it does not undermine the end-to-end benchmark results, which independently validate the trained latent space. Self-citations to CoD-Lite and SenseFlow are background/component choices supported by the paper's own ablations and measurements; they are not load-bearing circularity. No other prediction in the derivation chain is equivalent to a fitted parameter or training objective.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

No new physical entities, forces, or dimensions are introduced. 'Anchor-latent KL', 'NR-MMDiT', and 'Diffusion-NFT' are algorithms/architectures, not invented entities. The main load-bearing assumptions are the suitability of the FLUX.2-VAE anchor space and the validity of LLM-based evaluation.

free parameters (5)
  • CFG guidance scale w = 7.5
    Used in Delta_CA of the distillation gradient (Eq. 1). Hand-chosen; Turbo model output quality depends on it.
  • Adversarial guidance weight lambda_GAN = 0.13
    Set by hand in Eq. (1); Table 6 shows it changes Turbo benchmarks by up to +2.6 points on TIIF-Short, so the value is load-bearing.
  • Anchor-latent KL weight = 0.1
    Appendix A, Eq. (13)-(14). Controls how tightly Mage-VAE's latent matches FLUX.2-VAE; central to the compatibility claim made in Table 2.
  • Editing:generation data ratio in post-training/distillation = 4:1 (post-training), 3:1 (distillation)
    Section 5.2-5.3. Hand-chosen mixture; Table 7 shows effects are modest and benchmark-dependent.
  • Distillation set sizes = ~200K gen / ~250K edit samples
    Section 5.3. Curated by hand; no ablation on set size is reported.
axioms (5)
  • domain assumption FLUX.2-VAE latent space is a suitable anchor for generative training
    Sec 3.1.2. Mage-VAE is trained to reproduce patchified FLUX.2-VAE latents; if that space is suboptimal, downstream generation is bounded by it. Partially validated by Table 2 cross-tokenizer swap.
  • domain assumption Automated LLM/OCR judges produce valid quality rankings
    Sec 6.1 and Appendix D. GenEval/DPG/ImgEdit/GEdit/TextEdit scores all come from LLM judges or OCR; no human-correlation study is reported.
  • standard math Rectified flow matching is an adequate generative objective in Mage-VAE latent space
    Sec 3.2.2. Standard flow-matching loss; not in dispute.
  • domain assumption Native-resolution packing with per-sample 2D-RoPE and FlashAttention variable-length kernels preserves attention semantics
    Sec 3.2.1. No ablation compares packing against bucket-based training for quality; packed CFG is only validated via inference speedup (Table 3).
  • domain assumption Qwen3-VL text embeddings are sufficient conditioning
    Sec 3.2.2. Frozen Qwen3-VL-4B is used throughout; no comparison with other text encoders is reported.

pith-pipeline@v1.3.0-alltime-deepseek · 50860 in / 15388 out tokens · 132705 ms · 2026-08-01T13:32:25.426794+00:00 · methodology

0 comments
read the original abstract

Large-scale visual generators are increasingly capable but costly to train, fine-tune, and deploy. We introduce Mage-Flow, a compact 4B-scale generative stack for efficient text-to-image generation and instruction-based image editing. The stack is built from two co-designed components: Mage-VAE, a lightweight high-fidelity latent tokenizer, and a Native-Resolution Multimodal Diffusion Transformer trained with rectified flow matching. Mage-VAE uses one-step diffusion-style encoding and decoding with anchor-latent regularization, preserving the reconstruction quality of strong public VAEs while reducing tokenization cost by more than an order of magnitude. Together with native-resolution packing and stack-level CUDA kernel fusion, the stack supports flexible-resolution training and improves end-to-end training throughput by about $2.5\times$. Built on this foundation, we develop a complete model family with Base, RL-aligned, and Turbo variants for both generation and editing. Diffusion-NFT improves prompt following, text rendering, aesthetic quality, and editing fidelity, while few-step distillation with adversarial perceptual guidance produces 4-step Turbo models for low-latency inference. Despite its compact scale, Mage-Flow and Mage-Flow-Edit achieves competitive performance across standard generation and editing benchmarks. More importantly, the Turbo variants make high-resolution generation and editing practical for interactive use: at $1024^2$ resolution on a single NVIDIA A100 GPU, Mage-Flow-Turbo generates an image in 0.59s, and Mage-Flow-Edit-Turbo edits an image in 1.02s, while maintaining a small memory footprint. These results show that careful tokenizer--backbone--system co-design can deliver strong high-resolution generation and editing within an efficient 4B model family.

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model

    cs.CV 2026-07 conditional novelty 5.0

    Codec-guided sparse patch selection plus a lightweight speak/silent gate yields a 4B streaming VLM that is competitive on static tasks, stronger on video/spatial benchmarks, and much cheaper at inference.

Reference graph

Works this paper leans on

104 extracted references · 64 linked inside Pith · cited by 1 Pith paper

  1. [1]

    GPT-Image-1

    OpenAI. GPT-Image-1. https://openai.com/index/ introducing-4o-image-generation/, 2025

    2025

  2. [2]

    Nano banana pro

    Google. Nano banana pro. https://storage.googleapis.com/deepmind-media/ Model-Cards/Gemini-3-Pro-Image-Model-Card.pdf, 2025

    2025

  3. [3]

    Seedream 4.0: Toward next-generation multimodal image generation.arXiv preprint arXiv:2509.20427, 2025

    Team Seedream. Seedream 4.0: Toward next-generation multimodal image generation.arXiv preprint arXiv:2509.20427, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Z-Image: An efficient image generation foundation model with single-stream diffusion transformer

    Huanqia Cai, Sihan Cao, Ruoyi Du, Peng Gao, Steven Hoi, Zhaohui Hou, Shijie Huang, Dengyang Jiang, Xin Jin, Liangchen Li, Zhen Li, Zhong-Yu Li, David Liu, Dongyang Liu, Junhan Shi, Qilong Wu, Feng Yu, Chi Zhang, Shifeng Zhang, and Shilin Zhou. Z-Image: An efficient image generation foundation model with single-stream diffusion transformer. arXiv preprint ...

    Pith/arXiv arXiv 2025

  5. [5]

    Qwen-Image technical report.arXiv preprint arXiv:2508.02324, 2025

    Chenfei Wu, Jin Bai, Shuai Bai, Zhikai Chen, Kai Dang, Kaiyuan Gao, et al. Qwen-Image technical report.arXiv preprint arXiv:2508.02324, 2025

    Pith/arXiv arXiv 2025

  6. [6]

    FLUX.2: Frontier Visual Intelligence.https://bfl.ai/blog/flux-2, 2025

    Black Forest Labs. FLUX.2: Frontier Visual Intelligence.https://bfl.ai/blog/flux-2, 2025

    2025

  7. [7]

    Hun- yuanImage 3.0 technical report.arXiv preprint arXiv:2509.23951, 2025

    Siyu Cao, Hangting Chen, Peng Chen, Yiji Cheng, Yutao Cui, Xinchi Deng, et al. Hun- yuanImage 3.0 technical report.arXiv preprint arXiv:2509.23951, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    FireRed-Image-Edit-1.0 technical report.arXiv preprint arXiv:2602.13344, 2026

    Super Intelligence Team, Xiaohongshu. FireRed-Image-Edit-1.0 technical report.arXiv preprint arXiv:2602.13344, 2026

    arXiv 2026

  9. [9]

    CoD-Lite: Real-time diffusion-based generative image compression.arXiv preprint arXiv:2604.12525, 2026

    Zhaoyang Jia, Naifu Xue, Zihan Zheng, Jiahao Li, Bin Li, Xiaoyi Zhang, Zongyu Guo, Yuan Zhang, Houqiang Li, and Yan Lu. CoD-Lite: Real-time diffusion-based generative image compression.arXiv preprint arXiv:2604.12525, 2026

    Pith/arXiv arXiv 2026

  10. [10]

    Native-resolution image synthesis.arXiv preprint arXiv:2506.03131, 2025

    Zidong Wang, Lei Bai, Xiangyu Yue, Wanli Ouyang, and Yiyuan Zhang. Native-resolution image synthesis.arXiv preprint arXiv:2506.03131, 2025

    Pith/arXiv arXiv 2025

  11. [11]

    FlashAttention-2: Faster attention with better parallelism and work partitioning

    Tri Dao. FlashAttention-2: Faster attention with better parallelism and work partitioning. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  12. [12]

    Flashattention-4: Algorithm and kernel pipelining co-design for asymmetric hardware scaling

    Ted Zadouri, Markus Hoehnerbach, Jay Shah, Timmy Liu, Vijay Thakkar, and Tri Dao. Flashattention-4: Algorithm and kernel pipelining co-design for asymmetric hardware scaling. arXiv preprint arXiv:2603.05451, 2026

    arXiv 2026

  13. [13]

    Qwen3-VL technical report.arXiv preprint arXiv:2511.21631, 2025

    Qwen Team. Qwen3-VL technical report.arXiv preprint arXiv:2511.21631, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    DiffusionNFT: Online diffusion reinforcement with forward process.arXiv preprint arXiv:2509.16117, 2025

    Kaiwen Zheng, Huayu Chen, Haotian Ye, Haoxiang Wang, Qinsheng Zhang, Kai Jiang, Hang Su, Stefano Ermon, Jun Zhu, and Ming-Yu Liu. DiffusionNFT: Online diffusion reinforcement with forward process.arXiv preprint arXiv:2509.16117, 2025

    Pith/arXiv arXiv 2025

  15. [15]

    Decoupled DMD: CFG augmentation as the spear, distribution matching as the shield.arXiv preprint arXiv:2511.22677, 2025

    Dongyang Liu, Peng Gao, David Liu, Ruoyi Du, Zhen Li, Qilong Wu, Xin Jin, Sihan Cao, Shifeng Zhang, Hongsheng Li, and Steven Hoi. Decoupled DMD: CFG augmentation as the spear, distribution matching as the shield.arXiv preprint arXiv:2511.22677, 2025

    arXiv 2025

  16. [16]

    SenseFlow: Scaling distribution matching for flow-based text-to-image distillation.arXiv preprint arXiv:2506.00523, 2025

    Xingtong Ge, Xin Zhang, Tongda Xu, Yi Zhang, Xinjie Zhang, Yan Wang, and Jun Zhang. SenseFlow: Scaling distribution matching for flow-based text-to-image distillation.arXiv preprint arXiv:2506.00523, 2025. 52

    arXiv 2025

  17. [17]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Om- mer. High-resolution image synthesis with latent diffusion models. InCVPR, 2022. arXiv:2112.10752

    Pith/arXiv arXiv 2022

  18. [18]

    SDXL: Improving latent diffusion models for high-resolution image synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. InICLR, 2024. arXiv:2307.01952

    Pith/arXiv arXiv 2024

  19. [19]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. InICCV,

  20. [20]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. InICML, 2024. arXiv:2403.03206

    Pith/arXiv arXiv 2024

  21. [21]

    SANA: Efficient high-resolution image synthesis with linear diffusion transformers.arXiv preprint arXiv:2410.10629, 2025

    Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Haotian Tang, Yujun Lin, Zhekai Zhang, Muyang Li, Ligeng Zhu, Yao Lu, and Song Han. SANA: Efficient high-resolution image synthesis with linear diffusion transformers.arXiv preprint arXiv:2410.10629, 2025

    Pith/arXiv arXiv 2025

  22. [22]

    LongCat-Image technical report.arXiv preprint arXiv:2512.07584, 2025

    Meituan LongCat Team. LongCat-Image technical report.arXiv preprint arXiv:2512.07584, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    SDEdit: Guided image synthesis and editing with stochastic differential equations

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. InICLR, 2022. arXiv:2108.01073

    Pith/arXiv arXiv 2022

  24. [24]

    Prompt-to-prompt image editing with cross attention control

    Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. InICLR, 2023. arXiv:2208.01626

    Pith/arXiv arXiv 2023

  25. [25]

    Null-text inver- sion for editing real images using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inver- sion for editing real images using guided diffusion models. InCVPR, 2023. arXiv:2211.09794

    Pith/arXiv arXiv 2023

  26. [26]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. InICCV, 2023. arXiv:2302.05543

    Pith/arXiv arXiv 2023

  27. [27]

    IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models.arXiv preprint arXiv:2308.06721, 2023

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. IP-Adapter: Text compatible image prompt adapter for text-to-image diffusion models.arXiv preprint arXiv:2308.06721, 2023

    Pith/arXiv arXiv 2023

  28. [28]

    Tim Brooks, Aleksander Holynski, and Alexei A. Efros. InstructPix2Pix: Learning to follow image editing instructions. InCVPR, 2023. arXiv:2211.09800

    Pith/arXiv arXiv 2023

  29. [29]

    MagicBrush: A manually annotated dataset for instruction-guided image editing

    Kai Zhang, Lingbo Mo, Wenhu Chen, Huan Sun, and Yu Su. MagicBrush: A manually annotated dataset for instruction-guided image editing. InNeurIPS, 2023. arXiv:2306.10012

    Pith/arXiv arXiv 2023

  30. [30]

    AnyEdit: Mastering unified high-quality image editing for any idea.arXiv preprint arXiv:2411.15738, 2024

    Qifan Yu, Wei Chow, Zhongqi Yue, Kaihang Pan, Yang Wu, Xiaoyang Wan, Juncheng Li, Siliang Tang, Hanwang Zhang, and Yueting Zhuang. AnyEdit: Mastering unified high-quality image editing for any idea.arXiv preprint arXiv:2411.15738, 2024

    Pith/arXiv arXiv 2024

  31. [31]

    UltraEdit: Instruction-based fine-grained image editing at scale

    Haozhe Zhao, Xiaojian Ma, Liang Chen, Shuzheng Si, Rujie Wu, Kaikai An, Peiyu Yu, Minjia Zhang, Qing Li, and Baobao Chang. UltraEdit: Instruction-based fine-grained image editing at scale. InNeurIPS, 2024. arXiv:2407.05282

    Pith/arXiv arXiv 2024

  32. [32]

    OmniGen: Unified image generation.arXiv preprint arXiv:2409.11340, 2024

    Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. OmniGen: Unified image generation.arXiv preprint arXiv:2409.11340, 2024. 53

    Pith/arXiv arXiv 2024

  33. [33]

    FLUX.1 Kontext: Flow matching for in-context image generation and editing in latent space.arXiv preprint arXiv:2506.15742, 2025

    Black Forest Labs. FLUX.1 Kontext: Flow matching for in-context image generation and editing in latent space.arXiv preprint arXiv:2506.15742, 2025

    Pith/arXiv arXiv 2025

  34. [34]

    Step1X-Edit: A practical framework for general image editing.arXiv preprint arXiv:2504.17761, 2025

    Shiyu Liu et al. Step1X-Edit: A practical framework for general image editing.arXiv preprint arXiv:2504.17761, 2025

    Pith/arXiv arXiv 2025

  35. [35]

    In-context edit: Enabling instructional image editing with in-context generation in large scale diffusion transformer

    Zechuan Zhang, Ji Xie, Yu Lu, Zongxin Yang, and Yi Yang. In-context edit: Enabling instructional image editing with in-context generation in large scale diffusion transformer. arXiv preprint arXiv:2504.20690, 2025

    Pith/arXiv arXiv 2025

  36. [36]

    OmniGen2: Towards instruction-aligned multimodal generation.arXiv preprint arXiv:2506.18871, 2025

    Chenyuan Wu, Pengfei Zhou, Shitao Xiao, Jianpeng Zhang, Yong Tang, Zihan Huang, Yiyan Wang, et al. OmniGen2: Towards instruction-aligned multimodal generation.arXiv preprint arXiv:2506.18871, 2025

    Pith/arXiv arXiv 2025

  37. [37]

    UniWorld-V1: High-resolution semantic encoders for unified visual understanding and generation.arXiv preprint arXiv:2506.03147, 2025

    Bin Lin, Zongjian Li, Xinhua Cheng, Yuwei Niu, Yang Ye, Xianyi He, Shenghai Yuan, Wangbo Yu, Shaodong Wang, Yunyang Ge, Yatian Ning, and Li Yuan. UniWorld-V1: High-resolution semantic encoders for unified visual understanding and generation.arXiv preprint arXiv:2506.03147, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    DreamOmni2: Multimodal instruction-based editing and generation.arXiv preprint arXiv:2510.06679, 2025

    Bin Xia et al. DreamOmni2: Multimodal instruction-based editing and generation.arXiv preprint arXiv:2510.06679, 2025

    arXiv 2025

  39. [39]

    ChronoEdit: Towards temporal reasoning for image editing and world simulation.arXiv preprint arXiv:2510.04290, 2025

    Jiarui Wu et al. ChronoEdit: Towards temporal reasoning for image editing and world simulation.arXiv preprint arXiv:2510.04290, 2025

    arXiv 2025

  40. [41]

    Unified multimodal understanding and generation models: Advances, challenges, and opportunities

    Xinjie Zhang, Jintao Guo, Shanshan Zhao, Minghao Fu, Lunhao Duan, Jiakui Hu, Yong Xien Chng, Guo-Hua Wang, Qing-Guo Chen, Zhao Xu, Weihua Luo, and Kaifu Zhang. Unified multimodal understanding and generation models: Advances, challenges, and opportunities. arXiv preprint arXiv:2505.02567, 2025

    arXiv 2025

  41. [42]

    Janus-Pro: Unified multimodal understanding and generation with data and model scaling.arXiv preprint arXiv:2501.17811, 2025

    Xiaokang Chen, Zhiyu Wu, Xingchao Liu, Zizheng Pan, Wen Liu, Zhenda Xie, Xingkai Yu, and Chong Ruan. Janus-Pro: Unified multimodal understanding and generation with data and model scaling.arXiv preprint arXiv:2501.17811, 2025

    Pith/arXiv arXiv 2025

  42. [43]

    Transfusion: Predict the next token and diffuse images with one multi-modal model.arXiv preprint arXiv:2408.11039, 2024

    Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model.arXiv preprint arXiv:2408.11039, 2024

    Pith/arXiv arXiv 2024

  43. [44]

    Emerging properties in unified multimodal pretraining.arXiv preprint arXiv:2505.14683, 2025

    Chaorui Deng, Deyao Zhu, Kunchang Li, Chenhui Gou, Feng Li, Zeyu Wang, Shu Zhong, Weihao Yu, Xiaonan Nie, Ziang Song, Guang Shi, and Haoqi Fan. Emerging properties in unified multimodal pretraining.arXiv preprint arXiv:2505.14683, 2025

    Pith/arXiv arXiv 2025

  44. [45]

    Emu3.5: Native multimodal models are world learners.arXiv preprint arXiv:2510.26583, 2025

    Yuying Cui et al. Emu3.5: Native multimodal models are world learners.arXiv preprint arXiv:2510.26583, 2025

    Pith/arXiv arXiv 2025

  45. [46]

    Neural discrete representation learning

    Aaron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. InNeurIPS, 2017. arXiv:1711.00937. 54

    Pith/arXiv arXiv 2017

  46. [47]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. InCVPR, 2021. arXiv:2012.09841

    Pith/arXiv arXiv 2021

  47. [48]

    PSC: Posterior sampling-based compression

    Noam Elata, Tomer Michaeli, and Michael Elad. PSC: Posterior sampling-based compression. In15th International Conference on Sampling Theory and Applications, 2025

    2025

  48. [49]

    Idempotence and perceptual image compression

    Tongda Xu, Ziran Zhu, Dailan He, Yanghao Li, Lina Guo, Yuanyuan Wang, Zhe Wang, Hongwei Qin, Yan Wang, Jingjing Liu, and Ya-Qin Zhang. Idempotence and perceptual image compression. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  49. [50]

    Ultra lowrate image compression with semantic residual coding and compression-aware diffusion

    Anle Ke, Xu Zhang, Tong Chen, Ming Lu, Chao Zhou, Jiawen Gu, and Zhan Ma. Ultra lowrate image compression with semantic residual coding and compression-aware diffusion. arXiv preprint arXiv:2505.08281, 2025

    Pith/arXiv arXiv 2025

  50. [51]

    Oscar: One-step diffusion codec across multiple bit-rates.arXiv preprint arXiv:2505.16091, 2025

    Jinpei Guo, Yifei Ji, Zheng Chen, Kai Liu, Min Liu, Wang Rao, Wenbo Li, Yong Guo, and Yulun Zhang. Oscar: One-step diffusion codec across multiple bit-rates.arXiv preprint arXiv:2505.16091, 2025

    arXiv 2025

  51. [52]

    One-step diffusion- based image compression with semantic distillation

    Naifu Xue, Zhaoyang Jia, Jiahao Li, Bin Li, Yuan Zhang, and Yan Lu. One-step diffusion- based image compression with semantic distillation. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems, 2025

    2025

  52. [53]

    Stablecodec: Taming one-step diffusion for extreme image compression

    Tianyu Zhang, Xin Luo, Li Li, and Dong Liu. Stablecodec: Taming one-step diffusion for extreme image compression. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 17379–17389, October 2025

    2025

  53. [54]

    Diffusion model alignment using direct preference optimization

    Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. InCVPR, 2024. arXiv:2311.12908

    Pith/arXiv arXiv 2024

  54. [55]

    Using human feedback to fine-tune diffusion models without any reward model.arXiv preprint arXiv:2311.13231, 2023

    Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Qimai Li, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model.arXiv preprint arXiv:2311.13231, 2023

    Pith/arXiv arXiv 2023

  55. [56]

    Training diffusion models with reinforcement learning

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. InICLR, 2024. arXiv:2305.13301

    Pith/arXiv arXiv 2024

  56. [57]

    ImageReward: Learning and evaluating human preferences for text-to-image generation

    Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. ImageReward: Learning and evaluating human preferences for text-to-image generation. InNeurIPS, 2023. arXiv:2304.05977

    Pith/arXiv arXiv 2023

  57. [58]

    Flow-GRPO: Training flow matching models via online rl

    Jie Liu, Gongye Liu, Jiajun Liang, Yangguang Li, Jiaheng Liu, Xintao Wang, Pengfei Wan, Di Zhang, and Wanli Ouyang. Flow-GRPO: Training flow matching models via online rl. arXiv preprint arXiv:2505.05470, 2025

    Pith/arXiv arXiv 2025

  58. [59]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. arXiv:2010.02502

    Pith/arXiv arXiv 2021

  59. [60]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. InICLR, 2022. arXiv:2202.00512

    Pith/arXiv arXiv 2022

  60. [61]

    Consistency models

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. arXiv:2303.01469. 55

    Pith/arXiv arXiv 2023

  61. [62]

    Latent consistency models: Synthesizing high-resolution images with few-step inference.arXiv preprint arXiv:2310.04378, 2023

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. Latent consistency models: Synthesizing high-resolution images with few-step inference.arXiv preprint arXiv:2310.04378, 2023

    Pith/arXiv arXiv 2023

  62. [63]

    InstaFlow: One step is enough for high-quality diffusion-based text-to-image generation.arXiv preprint arXiv:2309.06380, 2024

    Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, and Qiang Liu. InstaFlow: One step is enough for high-quality diffusion-based text-to-image generation.arXiv preprint arXiv:2309.06380, 2024

    Pith/arXiv arXiv 2024

  63. [64]

    Freeman, and Taesung Park

    Tianwei Yin, Michaël Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T. Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In CVPR, 2024. arXiv:2311.18828

    Pith/arXiv arXiv 2024

  64. [65]

    Tianwei Yin, Michaël Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T. Freeman. Improved distribution matching distillation for fast image synthesis. InNeurIPS, 2024. arXiv:2405.14867

    Pith/arXiv arXiv 2024

  65. [66]

    Adversarial diffusion distillation.arXiv preprint arXiv:2311.17042, 2023

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation.arXiv preprint arXiv:2311.17042, 2023

    Pith/arXiv arXiv 2023

  66. [67]

    Hyper-SD: Trajectory segmented consistency model for efficient image synthesis.arXiv preprint arXiv:2404.13686, 2024

    Yuxi Ren, Xin Xia, Yanzuo Lu, Jiacheng Zhang, Jie Wu, Pan Xie, Xing Wang, and Xuefeng Xiao. Hyper-SD: Trajectory segmented consistency model for efficient image synthesis.arXiv preprint arXiv:2404.13686, 2024

    Pith/arXiv arXiv 2024

  67. [68]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InICML, 2021. arXiv:2103.00020

    Pith/arXiv arXiv 2021

  68. [69]

    DINOv2: Learning robust visual features without supervision.Transactions on Machine Learning Research (TMLR),

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jégou, Julien Mairal, Patrick Lab...

  69. [70]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873–12883, 2021

    2021

  70. [71]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Om- mer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  71. [72]

    Dico: Revitalizing convnets for scalable and efficient diffusion modeling.Advances in neural information processing systems, 38:123582–123618, 2026

    Yuang Ai, Qihang Fan, Xuefeng Hu, Zhenheng Yang, Ran He, and Huaibo Huang. Dico: Revitalizing convnets for scalable and efficient diffusion modeling.Advances in neural information processing systems, 38:123582–123618, 2026

    2026

  72. [73]

    Deco: Frequency- decoupled pixel diffusion for end-to-end image generation

    Zehong Ma, Longhui Wei, Shuai Wang, Shiliang Zhang, and Qi Tian. Deco: Frequency- decoupled pixel diffusion for end-to-end image generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 43600–43610, 2026

    2026

  73. [74]

    Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024

    Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models.arXiv preprint arXiv:2412.03603, 2024. 56

    Pith/arXiv arXiv 2024

  74. [75]

    Cod: A diffusion foundation model for image compression

    Zhaoyang Jia, Zihan Zheng, Naifu Xue, Jiahao Li, Bin Li, Zongyu Guo, Xiaoyi Zhang, Houqiang Li, and Yan Lu. Cod: A diffusion foundation model for image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 38420–38429, 2026

    2026

  75. [76]

    Projected gans converge faster.Advances in Neural Information Processing Systems, 34:17480–17492, 2021

    Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster.Advances in Neural Information Processing Systems, 34:17480–17492, 2021

    2021

  76. [77]

    Workshop and challenge on learned image compression(clic2020)

    Challenge on Learned Image Compression. Workshop and challenge on learned image compression(clic2020). CVPR2020Workshop, 2020. URL https://archive.compression. cc/

    2020

  77. [78]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401–4410, 2019

    2019

  78. [79]

    A self-supervised descriptor for image copy detection.Proc

    Ed Pizzi, Sreya Dutta Roy, Sugosh Nagavara Ravindra, Priya Goyal, and Matthijs Douze. A self-supervised descriptor for image copy detection.Proc. CVPR, 2022

    2022

  79. [80]

    The faiss library

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre- Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. The faiss library. 2024

    2024

  80. [81]

    Qwen3.5: Towards native multimodal agents, February 2026

    Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026. URLhttps: //qwen.ai/blog?id=qwen3.5

    2026

Showing first 80 references.