Split View: 이미지 생성·이해 기술 리포트, 무엇을 읽을 것인가 — 샘플 이미지 대신 설계를 읽기

Image Generation and Understanding Technical Reports: What to Read, and Why Design Beats Sample Images

• Introduction
• Why this article does not declare a winner
• What image reports are actually contesting
• Eleven technical reports worth reading
  • Scaling Rectified Flow Transformers for High-Resolution Image Synthesis
  • Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction
  • Emu3: Next-Token Prediction is All You Need
  • SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers
  • Janus-Pro: Unified Multimodal Understanding and Generation with Data and Model Scaling
  • FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space
  • Qwen-Image and Qwen-Image-2.0
  • Seedream 4.0: Toward Next-generation Multimodal Image Generation
  • Z-Image and Mage-Flow
  • InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models
  • Qwen-Image-Bench: From Generation to Creation in Text-to-Image Evaluation
• Limitations the authors state themselves
• What a practitioner should read first
• How and when this was verified
• Try it yourself
• References

Introduction

Part five is images. This is the domain where report reading gets distorted most easily. The sample images in a paper were chosen by the authors, and the comparisons people circulate online usually have no controlled prompt at all.

So this article reads design instead of samples. Tokenizer compression ratio, choice of text encoder, compute spent on training, how many sampling steps. Those values decide the constraints you will actually meet when you use the model.

Paper details were verified directly from the primary sources on 2026-08-12. This field moves fast, so check the current state yourself.

Why this article does not declare a winner

Image generation evaluation leans heavily on human preference studies, and results move with the prompt set and the rater pool. Automatic metrics only look at narrow axes. On top of that, recent reports often ship their own evaluation benchmark, which means numbers on that benchmark are by construction aligned with criteria the authors chose.

So this article never says one model is better. It records what each report made possible and what it gave up in exchange.

What image reports are actually contesting

[The axes image reports contest]
 Tokenizer  : compression ratio decides detail and training speed together
 Conditioning: what serves as the text encoder
 Unification: one model for generation, understanding and editing
 Step count : how few sampling steps are enough
 Text       : can it render legible characters inside the image

The last line is the axis that suddenly grew in recent reports, because they started targeting outputs where text is the point: slides, posters, infographics.

Eleven technical reports worth reading

Scaling Rectified Flow Transformers for High-Resolution Image Synthesis

arxiv.org/abs/2403.03206 — submitted 2024-03-05.

Proposes rectified flow in place of diffusion, together with noise sampling that weights perceptually relevant scales. The transformer with separate weights for the two modalities and bidirectional information flow becomes the starting point for many later reports. It also contains the scaling observation that validation loss and synthesis quality track each other predictably.

Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction

arxiv.org/abs/2404.02905 — submitted 2024-04-03, v2 on 2024-06-10.

Redefines autoregressive image generation from raster-scan next-token prediction to coarse-to-fine next-scale prediction. On ImageNet 256 the authors report FID moving from 18.65 to 1.73 and IS from 80.4 to 350.2. They describe the work as having initially emulated the properties observed in LLMs, which is their own signal that these are early-stage results.

Emu3: Next-Token Prediction is All You Need

arxiv.org/abs/2409.18869 — submitted 2024-09-27.

Turns images, text and video all into discrete tokens and handles them with next-token prediction alone, using neither diffusion nor compositional frameworks. Read it as the representative case for simplifying multimodal design rather than as a performance claim.

SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers

arxiv.org/abs/2410.10629 — submitted 2024-10-14, v3 on 2024-10-20.

A report with efficiency as the stated goal. The core moves are a deep compression autoencoder at 32x instead of the usual 8x, linear attention replacing standard attention in the diffusion transformer, and a decoder-only LLM as text encoder in place of T5. The authors report generating a 1024x1024 image in under one second on a 16GB laptop GPU.

Janus-Pro: Unified Multimodal Understanding and Generation with Data and Model Scaling

arxiv.org/abs/2501.17811 — submitted 2025-01-29.

Reports advances in both understanding and text-to-image instruction following through an optimized training strategy, expanded data and larger model size, along with improved stability in text-to-image generation. Note that the arXiv page carries an administrative note about text overlap with an earlier paper, so reading the two together is the accurate approach.

FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space

arxiv.org/abs/2506.15742 — Black Forest Labs, submitted 2025-06-17, v2 on 2025-06-24.

Merges generation and editing into one flow matching model, using sequence concatenation to carry semantic context from both text and image inputs. The problem the authors name is precise: current editing models degrade in character consistency and stability across multiple turns. The accompanying KontextBench holds 1,026 image-prompt pairs across five task categories.

Qwen-Image and Qwen-Image-2.0

arxiv.org/abs/2508.02324 (submitted 2025-08-04) and arxiv.org/abs/2605.10730 (submitted 2026-05-11).

The first report takes on text rendering inside images directly, with a pipeline spanning collection, filtering and synthesis, plus progressive training from simple to complex text, targeting both alphabetic and logographic writing systems. The second extends toward text-rich outputs like slides, posters and infographics with instructions of up to 1K tokens. Qwen-Image-VAE-2.0 (arxiv.org/abs/2605.13565), which treats the tokenizer separately, belongs to the same line.

Seedream 4.0: Toward Next-generation Multimodal Image Generation

arxiv.org/abs/2509.20427 — Team Seedream, submitted 2025-09-24, v3 on 2025-12-10.

Puts text-to-image, image editing and multi-image composition in one framework. A VAE that reduces image token counts enables native resolutions from 1K to 4K, and the authors report up to 1.8 seconds for a 2K image.

Z-Image and Mage-Flow

arxiv.org/abs/2511.22699 (submitted 2025-11-27, v5 on 2026-07-06) and arxiv.org/abs/2607.19064 (submitted 2026-07-21, v2 on 2026-07-22).

Both reports push against the scale-up current. The first is a 6-billion-parameter single-stream diffusion transformer trained in 314,000 H800 GPU hours, whose distilled variant is stated to run sub-second inference on consumer devices with less than 16GB VRAM. The second is 4 billion parameters, pairing a lightweight tokenizer that uses one-step diffusion encoding with a rectified flow transformer, reporting 0.59 seconds for 1024x1024 generation and 1.02 seconds for editing on an A100.

InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models

arxiv.org/abs/2504.10479 — submitted 2025-04-14, v3 on 2025-04-19.

The understanding-side entry. It proposes native multimodal pre-training that develops visual and language ability jointly from the start rather than aligning them afterward. The 78-billion-parameter variant reports 72.2 on MMMU, an author-run measurement.

Qwen-Image-Bench: From Generation to Creation in Text-to-Image Evaluation

arxiv.org/abs/2605.28091 — submitted 2026-05-27, v2 on 2026-06-25.

The evaluation entry, and possibly the one on this list that stays useful longest. It starts from the concern that existing benchmarks remain anchored in foundational criteria and miss the capabilities that matter in authentic creative practice. It decomposes 1,000 stratified prompts into five pillars, 23 sub-capabilities and 56 rubrics, producing granular diagnostics rather than one opaque score.

Limitations the authors state themselves

• The VAR authors describe their results as having initially emulated LLM properties, signalling early-stage findings.
• The Janus-Pro arXiv page carries an administrative note about text overlap with a prior paper.
• The multi-turn consistency degradation FLUX.1 Kontext identifies remains an open problem for this whole family.
• The Qwen-Image-Bench authors concede that existing benchmarks fail to capture the capabilities that matter in real creative work. Keep that in mind when reading numbers from reports that ship their own benchmark.
• Every performance statement above is an author-run measurement, and many are values on a benchmark the authors built themselves.

What a practitioner should read first

If you want the principles first, the rectified flow transformer report is the starting point. Most recent reports are variations on that structure.

If cost is the constraint, read SANA, Z-Image and Mage-Flow as a group. All three look for the answer in the tokenizer and the sampling steps rather than in scale.

If you are building an editing workflow, FLUX.1 Kontext. Its problem definition of multi-turn consistency goes straight into product design.

If text inside images matters, the Qwen-Image line. Not many reports address that axis explicitly yet.

If you need to build evaluation, borrow the decomposition approach from Qwen-Image-Bench. Moving from one score to per-capability diagnostics fits practice better.

How and when this was verified

Every report cited here was verified on 2026-08-12 by opening the arXiv abstract page directly. Candidates that could not be opened were not cited. Every number quoted is as reported by the authors.

Try it yourself

• Neural Network Architecture Explorer — following the diffusion transformer variants requires the basic block structure to be in your hands.
• ML Training Data Explorer — see by example what data image generation models consume.

Previous in the series: Speech Recognition and Synthesis Technical Reports: What to Read

Next in the series: How to Read Leaderboards and Benchmarks

References

• Scaling Rectified Flow Transformers for High-Resolution Image Synthesis (arXiv 2403.03206): arxiv.org/abs/2403.03206
• Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction (arXiv 2404.02905): arxiv.org/abs/2404.02905
• Emu3: Next-Token Prediction is All You Need (arXiv 2409.18869): arxiv.org/abs/2409.18869
• SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformers (arXiv 2410.10629): arxiv.org/abs/2410.10629
• Janus-Pro: Unified Multimodal Understanding and Generation with Data and Model Scaling (arXiv 2501.17811): arxiv.org/abs/2501.17811
• InternVL3: Exploring Advanced Training and Test-Time Recipes for Open-Source Multimodal Models (arXiv 2504.10479): arxiv.org/abs/2504.10479
• FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space (arXiv 2506.15742): arxiv.org/abs/2506.15742
• Qwen-Image Technical Report (arXiv 2508.02324): arxiv.org/abs/2508.02324
• Seedream 4.0: Toward Next-generation Multimodal Image Generation (arXiv 2509.20427): arxiv.org/abs/2509.20427
• Z-Image: An Efficient Image Generation Foundation Model with Single-Stream Diffusion Transformer (arXiv 2511.22699): arxiv.org/abs/2511.22699
• Qwen-Image-2.0 Technical Report (arXiv 2605.10730): arxiv.org/abs/2605.10730
• Qwen-Image-VAE-2.0 Technical Report (arXiv 2605.13565): arxiv.org/abs/2605.13565
• Qwen-Image-Bench: From Generation to Creation in Text-to-Image Evaluation (arXiv 2605.28091): arxiv.org/abs/2605.28091
• Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing (arXiv 2607.19064): arxiv.org/abs/2607.19064