mlsanityData to 7 Oct 2026

STARFlow2: Bridging Language Models and Normalizing Flows for Unified Multimodal Generation

arXiv 2605.08029 ↗First posted 8 May 2026cs.CV cs.LGcited by 2

On the map: Efficient LLM inference and serving → Visual token pruning in multimodal models

Abstract (arXiv)

Deep generative models have advanced rapidly across text and vision, motivating unified multimodal systems that can understand, reason over, and generate interleaved text-image sequences. Most existing approaches combine autoregressive language modeling with diffusion-based image generators, inheriting a structural mismatch between causal text generation and iterative visual denoising. We observe that autoregressive normalizing flows are autoregressive Transformers--sharing the same causal mask, KV-cache mechanism, and left-to-right structure as LLMs--making them the most natural paradigm for true unified multimodal generation. We present STARFlow2, built on the Pretzel architecture that vertically interleaves a pretrained VLM stream with a TarFlow stream via residual skip connections, both operating under the same causal mask. Combined with a deep-shallow flow design and a unified FAE latent space, STARFlow2 enables cache-friendly interleaved generation where both text and visual outputs directly enter the KV-cache without re-encoding. Experiments demonstrate strong performance across image generation and multimodal understanding benchmarks, validating autoregressive flows as a viable foundation for unified multimodal modeling.

Key facts. Cited by 2 papers in the map (data through 7 Oct 2026). Placed in Visual token pruning in multimodal models, one of 12 directions in Efficient LLM inference and serving. Builds on 63 of its 66 references that are in the map. First posted on arXiv 8 May 2026.

Builds on

63 of 66 references.

References · 63

  1. 2020Denoising Diffusion Probabilistic Modelscited by 20,851
  2. 2021High-Resolution Image Synthesis with Latent Diffusion Modelscited by 18,694
  3. 2023GPT-4 Technical Reportcited by 18,132
  4. 2023DINOv2: Learning Robust Visual Features without Supervisioncited by 7,764
  5. 2022Scalable Diffusion Models with Transformerscited by 5,931
  6. 2022Flow Matching for Generative Modelingcited by 5,426
  7. 2025Qwen2.5-VL Technical Reportcited by 5,418
  8. 2017Neural Discrete Representation Learningcited by 4,650
  9. 2023Improved Baselines with Visual Instruction Tuningcited by 4,362
  10. 2024Scaling Rectified Flow Transformers for High-Resolution Image Synthesiscited by 4,094
  11. 2017FiLM: Visual Reasoning with a General Conditioning Layercited by 2,286
  12. 2015Variational Inference with Normalizing Flowscited by 2,161
38 more
  1. 2023MMBench: Is Your Multi-modal Model an All-around Player?cited by 2,105
  2. 2023MMMU: A Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark for Expert AGIcited by 2,044
  3. 2016Density estimation using Real NVPcited by 2,029
  4. 2023MME: A Comprehensive Evaluation Benchmark for Multimodal Large Language Modelscited by 1,624
  5. 2018Glow: Generative Flow with Invertible 1x1 Convolutionscited by 1,539
  6. 2021Conceptual 12M: Pushing Web-Scale Image-Text Pre-Training To Recognize Long-Tail Visual Conceptscited by 1,209
  7. 2014NICE: Non-linear Independent Components Estimationcited by 1,052
  8. 2025Qwen-Image Technical Reportcited by 973
  9. 2025FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Spacecited by 960
  10. 2023GenEval: An Object-Focused Framework for Evaluating Text-to-Image Alignmentcited by 946
  11. 2023SEED-Bench: Benchmarking Multimodal LLMs with Generative Comprehensioncited by 929
  12. 2025Emerging Properties in Unified Multimodal Pretrainingcited by 855
  13. 2024Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMscited by 822
  14. 2016A Diagram Is Worth A Dozen Imagescited by 793
  15. 2025Janus-Pro: Unified Multimodal Understanding and Generation with Data and Model Scalingcited by 787
  16. 2024Emu3: Next-Token Prediction is All You Needcited by 761
  17. 2024Show-o: One Single Transformer to Unify Multimodal Understanding and Generationcited by 734
  18. 2017Masked Autoregressive Flow for Density Estimationcited by 680
  19. 2016Improving Variational Inference with Inverse Autoregressive Flowcited by 636
  20. 2023Language Model Beats Diffusion -- Tokenizer is Key to Visual Generationcited by 633
  21. 2024Transfusion: Predict the Next Token and Diffuse Images with One Multi-Modal Modelcited by 468
  22. 2024ELLA: Equip Diffusion Models with LLM for Enhanced Semantic Alignmentcited by 455
  23. 2025BLIP3-o: A Family of Fully Open Unified Multimodal Models-Architecture, Training and Datasetcited by 363
  24. 2024Visual Sketchpad: Sketching as a Visual Chain of Thought for Multimodal Language Modelscited by 260
  25. 2019Flow++: Improving Flow-Based Generative Models with Variational Dequantization and Architecture Designcited by 258
  26. 2025Show-o2: Improved Native Unified Multimodal Modelscited by 242
  27. 2025Transfer between Modalities with MetaQueriescited by 231
  28. 2024MetaMorph: Multimodal Understanding and Generation via Instruction Tuningcited by 205
  29. 2023JourneyDB: A Benchmark for Generative Image Understandingcited by 203
  30. 2024TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generationcited by 200
  31. 2024Open-MAGVIT2: An Open-Source Project Toward Democratizing Auto-regressive Visual Generationcited by 153
  32. 2024Mixture-of-Transformers: A Sparse and Scalable Architecture for Multi-Modal Foundation Modelscited by 151
  33. 2024OmniEdit: Building Image Editing Generalist Models Through Specialist Supervisioncited by 145
  34. 2024LMFusion: Adapting Pretrained Language Models for Multimodal Generationcited by 120
  35. 2025Mogao: An Omni Foundation Model for Interleaved Multi-Modal Generationcited by 108
  36. 2024SEED-Data-Edit Technical Report: A Hybrid Dataset for Instructional Image Editingcited by 97
  37. 2024Normalizing Flows are Capable Generative Modelscited by 89
  38. 2025Zebra-CoT: A Dataset for Interleaved Vision Language Reasoningcited by 78

50 most cited of 63 All 63 (JSON)

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2026: 2.

2 papers.

  1. 2026Latent Reasoning with Normalizing Flowscited by 4
  2. 2026MIMFlow: Integrating Masked Image Modeling with Normalizing Flows for End-to-End Image Generationcited by 1

Similar, not cited

Close in wording, no citation either way

  1. 2026Multimodal Flow: Unified Flow Modeling of Language and Vision in Embedding Spaces
  2. 2025Muddit: Liberating Generation Beyond Text-to-Image with a Unified Discrete Diffusion Model
  3. 2024JanusFlow: Harmonizing Autoregression and Rectified Flow for Unified Multimodal Understanding and Generation
  4. 2024JetFormer: An Autoregressive Generative Model of Raw Images and Text
  5. 2026Unified Text-Image Generation with Weakness-Targeted Post-Training
  6. 2025Growing Visual Generative Capacity for Pre-Trained MLLMs
  7. 2026TextLDM: Language Modeling with Continuous Latent Diffusion
  8. 2024ANOLE: An Open, Autoregressive, Native Large Multimodal Models for Interleaved Image-Text Generation