PaperlayerView exact revision on arXiv

Source-linked AI summary

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, 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

arXiv:2607.19064v2cs.CVcs.AIcs.LGcs.MMeess.IV

TL;DR

Practical, reproducible image generation and editing remain difficult under realistic compute budgets as capable systems grow larger. Mage-Flow addresses this with a co-designed, efficient 4B-scale stack that achieves competitive benchmark performance with fast inference and low memory use.

  • Problem

    Growing model scale raises the cost of inference, fine-tuning, ablations, and deployment, leaving practical open visual generation and editing under realistic compute budgets insufficiently addressed.

  • Method

    Mage-Flow co-designs a lightweight high-fidelity tokenizer, native-resolution diffusion transformer, and fused-kernel infrastructure in a compact 4B-scale generation-and-editing stack.

  • Results

    Competitive performance across standard generation and editing benchmarks is achieved against substantially larger open-source systems while maintaining low memory usage and fast inference.

  • Takeaways & Limitations

    A 4B-scale stack can serve as a practical, research-friendly foundation for generation, controllable editing, alignment, and vertical-domain applications without tens-of-billions-parameter backbones.

  • Takeaways & Limitations

    Four-step editing gains on general benchmarks are benchmark-dependent rather than uniform.

Abstract

from arXiv · show

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.

1 Introduction

Mage-Flow addresses the gap between strong but compute-intensive visual generators and practical open research by co-designing a compact 4B stack for text-to-image generation and instruction-based editing. Its lightweight tokenizer, native-resolution backbone, and fused-kernel infrastructure support competitive quality with efficient training, inference, and memory use.

  • Motivation: Open visual generation remains difficult to study and deploy under realistic compute budgets as competitive systems increasingly rely on large backbones.The introduction contrasts limited transparency in closed systems with open generators ranging from 6B to 80B parameters.
  • Stack design: Mage-Flow combines Mage-VAE with a 4B Native-Resolution Multimodal Diffusion Transformer trained by rectified flow matching for generation and editing.The shared stack operates in Mage-VAE latent space and uses the same backbone for text-to-image generation and instruction-based editing.
  • Systems efficiency: MFU increases from approximately 14% to 29%, while fused CUDA kernels achieve about a 2.5× end-to-end training speedup.The optimization fuses memory-bound operator chains across Mage-VAE, the Qwen3-VL text encoder, and NR-MMDiT.
  • Model family and results: At 1024^2 resolution on one NVIDIA A100, Mage-Flow-Turbo generates images in 0.59s and Mage-Flow-Edit-Turbo edits images in 1.02s.The family includes Base, RL-aligned, and Turbo variants, with Turbo models produced through few-step distillation and adversarial perceptual guidance.
  • Model family and results: Peak GPU memory remains around 18-20 GB, while the models deliver competitive or superior benchmark performance against much larger open-source systems.The compared systems include Qwen-Image, Z-Image, FLUX.2, and FireRed-Image-Edit.
  • Tokenizer: Mage-VAE uses one-step diffusion-style encoding and decoding with anchor-latent KL regularization to preserve strong public VAE reconstruction quality while reducing high-resolution tokenization cost.The tokenizer is redesigned because it runs during training, inference, and repeated editing, with cost increasing rapidly at higher resolutions.

2 Related Work

Related work spans the evolution of text-to-image generation, instruction-based editing, unified multimodal models, visual tokenization, preference alignment, and few-step diffusion distillation. Mage-Flow is positioned as a specialized, efficient 4B MMDiT system within these directions.

  • Text-to-image generators: Text-to-image generation has progressed from pixel-space diffusion and U-Net latent diffusion toward diffusion transformers and rectified-flow models.Stable Diffusion popularized compressed VAE latents, SDXL scaled U-Net conditioning and aspect-ratio training, and Diffusion Transformers shifted the field toward transformer backbones.
  • Instruction-based image editors: Instruction-based editing evolved from inversion and guidance methods requiring per-image optimization toward end-to-end instruction-tuned editors.SDEdit enabled global edits, while Prompt-to-Prompt and Null-text Inversion improved localized editing through cross-attention manipulation or optimized embeddings.
  • Unified multimodal generation models: Unified multimodal models combine image understanding, generation, and editing in one backbone, but are typically larger and architecturally distinct from MMDiT-based generators.Janus-Pro, Transfusion, BAGEL, and Emu3.5 represent distinct approaches to unified multimodal modeling, whereas Mage-Flow specializes a 4B MMDiT stack.
  • Visual tokenization: Visual tokenization defines a major efficiency frontier, with continuous-latent VAEs dominant in diffusion systems and diffusion codecs trading perceptual quality against latency and backbone cost.Multi-step diffusion codecs offer strong perceptual quality at low bitrates but are slow, while one-step codecs improve latency yet often use heavy DiT- or U-Net-style backbones.
  • Post-training and few-step distillation: Post-training and distillation align diffusion or flow-matching generators with preferences while reducing deployment sampling steps through preference optimization, reinforcement learning, solver improvements, and few-step distillation.Diffusion-DPO, D3PO, DDPO, and Flow-GRPO address alignment, while progressive distillation, consistency methods, DMD-family approaches, and adversarial guidance reduce step counts; Mage-Flow uses Decoupled DMD with VFM-based adversarial guidance for 4-step Turbo variants.

3 Mage-Flow Stack · 3.1 Mage-VAE: Lightweight Image Tokenizer · 3.2 Native-Resolution MMDiT

Mage-Flow combines a lightweight high-fidelity Mage-VAE tokenizer with a 4B native-resolution multimodal diffusion transformer trained by rectified flow matching. Anchor-latent supervision preserves reconstruction and downstream compatibility while native-resolution packing supports heterogeneous resolutions and efficient inference.

  • 3 Mage-Flow Stack: Mage-Flow integrates Mage-VAE with a 4B Native-Resolution Multimodal Diffusion Transformer trained with rectified flow matching.Mage-VAE supplies a compact generation-ready latent space, while NR-MMDiT processes packed latent sequences with flexible resolutions and aspect ratios.
  • 3.1.1 Architecture: Mage-VAE uses fully convolutional one-step diffusion-style encoding and decoding to avoid global attention and keep decoding cost nearly linear in image resolution.The decoder reconstructs RGB pixels directly from latents, while the encoder is designed as its architectural dual for efficient image-to-latent conversion.
  • 3.1.2 KL regularization with an anchor latent distribution: Mage-VAE replaces standard-Gaussian matching with KL regularization toward an anchor latent distribution induced by FLUX.2-VAE.The anchor follows FLUX.2-VAE’s 32 latent channels and 8× spatial reduction, with its latents typically 2× patchified before diffusion-transformer input.
  • 3.1.3 Training: Mage-VAE training has three stages: separate multi-step flow-matching pretraining, one-step decoder distillation, and subsequent tokenizer training.Stage I trains the encoder and decoder separately; Stage II distills the decoder using reconstruction and DMD losses.
  • 3.1.4 Evaluation: Mage-VAE preserves generation-ready latent structure: swapping it with FLUX.2-VAE while keeping downstream backbones fixed produces similar or comparable generation and editing benchmark performance.Anchor-latent supervision enables efficient VAE distillation without breaking reconstruction quality or cross-generator compatibility.
  • 3.2.1 Model Architecture and Native Packing: Native-resolution packing removes fixed resolution buckets, allowing each update to mix heterogeneous image sizes and aspect ratios under a fixed token budget.This preserves native latent grids, improves batching flexibility, and turns resolution diversity into a training signal.
  • 3.2.1 Model Architecture and Native Packing: 1.09×–1.15×: packed classifier-free guidance accelerates inference across Mage-Flow and Mage-Edit variants while preserving the original denoising trajectory.Conditional and unconditional branches are evaluated in a single forward pass, avoiding redundant separate CFG computation.

3.3 Training Infrastructure

Mage-Flow improves training efficiency through co-design of lightweight tokenization and stack-level kernel fusion. Fused kernels reduce memory-bound overhead across repeated VAE, text-encoder, and NR-MMDiT blocks, while the full system raises MFU from 13.88% to 29.28%.

  • Kernel Fusion: Kernel fusion targets memory-bound operator chains across Mage-VAE, Qwen3-VL, and NR-MMDiT repeated blocks.Fused operations include normalization, activation, residual updates, adaptive normalization, rotary embeddings, and gated residual updates.
  • Throughput Results: 1.41× speedup results from replacing FLUX.2-VAE with Mage-VAE, reducing per-step time from 1.9285 s to 1.3647 s.This result demonstrates the training benefit of a lightweight tokenizer before additional fusion gains.
  • Throughput Results: 29.28% MFU, up from 13.88%, is achieved by the full system through tokenizer replacement and stack-level kernel fusion.Fusing the repeated NR-MMDiT blocks provides the largest improvement because the 4B diffusion backbone dominates each training step.
  • Design Principle: Efficient native-resolution generation requires co-designing the model and training system to reduce arithmetic cost and memory-bound overhead.Lightweight tokenization lowers arithmetic cost, while kernel fusion addresses repeated VAE, text-encoder, and MMDiT blocks.

4 Data Collection and Curation

Mage-Flow uses separate curation pipelines for text-to-image generation and instruction-based editing, targeting visual quality, safety, semantic alignment, diversity, and capability coverage. Generation data undergoes staged filtering, deduplication, multi-level captioning, and concept-aware supplementation, while editing data is constructed and filtered as source–instruction–target triples.

  • Data pipelines: Two complementary pipelines curate image–text pairs for generation and source-image, instruction, and target-image triples for editing.Both pipelines are designed to improve visual quality, safety, semantic alignment, diversity, and coverage of capability-specific skills.
  • Generation data: Roughly 10B raw image–text pairs feed a four-stage generation curation pipeline: sample-level filtering, cross-sample deduplication, multi-granularity captioning, and concept-aware synthesis.The stages remove unsuitable samples and redundancy, standardize textual supervision, and supplement capability-critical long-tail concepts.
  • Generation data: Sample-level filters remove corrupted, low-resolution, malformed, visually poor, unsafe, or unsuitable image–caption pairs before further curation.File-information checks cover corruption, resolution, pixel count, aspect ratio, and orientation; image-content filters assess decoded-image quality and safety.
  • Generation data: SSCD descriptors and FAISS nearest-neighbor search deduplicate retained images within and across datasets, grouping pairs with cosine similarity above 0.9 as duplicates.The descriptor is robust to reencoding, cropping, resizing, and light edits.
  • Editing data: Editing data combines open-source and in-house triples, then applies VLM-based majority-vote filtering, edit-type tagging, and category balancing.Three independent Qwen3.5-9B experts evaluate source image, target image, and instruction consistency using different prompts and partially overlapping criteria.

5 Training

Mage-Flow uses a unified latent-space and backbone recipe for generation and editing, with task-specific conditioning and data mixtures across progressive pre-training, supervised fine-tuning, Diffusion-NFT alignment, and four-step distillation. Editing adaptation preserves open-ended synthesis while supporting unified semantic, appearance, restoration, and structure-aware edits.

  • Unified recipe: Both tasks share Mage-VAE latents, the 4B Native-Resolution MMDiT backbone, and rectified-flow training, differing in conditioning format and data mixture.Generation uses prompts, while editing uses instructions with source image(s); the recipe proceeds through progressive pre-training and supervised fine-tuning.
  • Text-to-image generation: 1.2B image–text pairs at 256 × 256 initialize broad visual–language alignment, followed by 150M curated high-quality examples at 1024-pixel native-aspect-ratio resolution for Mage-Flow-Base.The curriculum progressively increases resolution, data quality, and concept-aware reweighting before supervised fine-tuning with stricter quality and content filters.
  • Instruction-based editing: The editing checkpoint supports a unified one-to-many interface for semantic edits, appearance transformations, restoration, and structure-aware outputs from the same source image.This capability breadth is learned through instruction-based adaptation rather than separate task-specific models.
  • Alignment and distillation: 4-step Turbo models are distilled from RL-aligned checkpoints, with adversarial perceptual guidance preserving generation and text-editing quality while general editing gains remain benchmark-dependent.The distillation produces Mage-Flow-Turbo and Mage-Flow-Edit-Turbo and reduces inference to a four-step rectified-flow trajectory.

6 Experiments

Across text-to-image and instruction-based editing benchmarks, the 4B Mage-Flow family achieves competitive or leading quality against substantially larger systems. Its Turbo variants preserve strong generation and editing behavior with only four denoising steps, while editing results remain strongest on several evaluated metrics.

  • Text-to-image generation: With 4B parameters, Mage-Flow achieves the best GenEval score and competitive performance on DPG-Bench, TIIF-Bench, OneIG, and LongText.It also obtains one of the strongest open-source CVTG-2K results, approaching the best 32B open-source specialist model.
  • Prompt following and compositional understanding: Mage-Flow achieves the strongest overall GenEval score and remains competitive on DPG-Bench, outperforming most unified baselines of similar or larger scale.Its strongest GenEval categories include single-object generation, two-object generation, counting, and positional relations.
  • Text rendering and bilingual generation: Mage-Flow delivers near-best open-source multi-region text-rendering accuracy on CVTG-2K and strong English long-text rendering on LongText.Chinese LongText performance remains competitive but leaves room for further data supplementation.
  • Text rendering and bilingual generation: Mage-Flow-Base, Mage-Flow, and Mage-Flow-Turbo maintain strong alignment and text-rendering ability on both English and Chinese OneIG splits.The results indicate that the multi-granularity captioning and bilingual data pipeline transfer to fine-grained generation.
  • Instruction-based image editing: Mage-Flow-Edit reaches ImgEdit-Bench overall scores of 4.28, 4.34, and 4.38 for Base, RL-aligned, and Turbo, respectively.The variants outperform earlier instruction-tuned models and remain competitive with recent MMDiT-based editing systems.
  • GEdit-Bench: On GEdit-Bench, Mage-Flow-Edit scores 8.127 on English and 8.123 on Chinese, while Turbo reaches 8.271 and 8.264 with four denoising steps.Both variants outperform Qwen-Image-Edit-2511, FireRed-Image-Edit-1.0, and LongCat-Image-Edit under the same protocol.

7 Conclusion

Mage-Flow is a compact, efficient 4B-scale stack for text-to-image generation and instruction-based image editing, jointly designed across tokenizer, native-resolution backbone, and training infrastructure. Its Base, RL-aligned, and Turbo variants achieve competitive benchmark performance while supporting practical generation, editing, alignment, and domain applications.

  • Contributions: Mage-Flow presents a compact, efficient 4B-scale framework for text-to-image generation and instruction-based image editing.Its design jointly integrates a lightweight latent tokenizer, a native-resolution diffusion Transformer, and system-level training infrastructure.
  • System design: Mage-VAE reduces latent encoding and decoding cost while preserving reconstruction fidelity, and NR-MMDiT supports training and inference across resolutions and aspect ratios.
  • Model family: Mage-Flow and Mage-Flow-Edit include Base, RL-aligned, and Turbo variants, with Diffusion-NFT improving alignment and Decoupled-DMD producing 4-step Turbo models for low-latency inference.The reported alignment targets include prompt-following, text-rendering, aesthetic, and editing preferences.
  • Results: The resulting models achieve competitive performance across standard generation and editing benchmarks.
  • Implications: A 4B-scale stack can serve as a practical and research-friendly foundation for image generation, controllable editing, post-training alignment, and vertical-domain applications.The conclusion attributes this capability to an efficient tokenizer, native-resolution modeling, and carefully optimized training infrastructure.

Contributor List

This section lists the paper’s contributors, beginning with Xinjie Zhang and including 23 additional named contributors.

  • The contributor list begins with Xinjie Zhang, Peng Zhang, Shicheng Zheng, Jinghao Guo, Zhaoyang Jia, and Yifei Shen.
  • Additional contributors are Xun Guo, Yuxuan Luo, Jiahao Li, Wenxuan Xie, Fanyi Pu, Xiaoyi Zhang, Kaichen Zhang, and Zongyu Guo.
  • The remaining named contributors are Tianci Bi, Dongnan Gui, Zhening Liu, Zimo Wen, Zihan Zheng, Senqiao Yang, Xiao Li, Jinglu Wang, Bin Li, and Yan Lu.

A Detailed Mage-VAE Training Process

Mage-VAE is trained in three stages: multi-step flow-matching pre-training, one-step decoder distillation, and joint one-step VAE optimization. The resulting model supports one-step encoding and decoding with high-fidelity reconstruction and a compact, regularized latent space.

  • Stage I: Multi-step flow-matching pre-training: Stage I pre-trains the encoder and decoder as multi-step diffusion models with flow matching in the X-prediction parameterization.The decoder reconstructs image pixels from anchor latents, while the encoder generates patchified FLUX.2-VAE anchor latents from image pixels.
  • Stage II: One-step decoder distillation: Stage II distills the multi-step decoder into a one-step decoder that reconstructs images from anchor latents with a single network evaluation.Its objective combines pixel-level and perceptual reconstruction losses with a DINOv2-projected GAN loss and DMD loss.
  • Stage III: Joint one-step VAE optimization: Stage III converts the encoder into a one-step model and jointly optimizes it with the distilled decoder while regularizing the learned latent distribution toward the frozen anchor-latent distribution.The KL regularizer preserves the pretrained anchor-latent structure and prevents distribution drift during perceptual fine-tuning.
  • Stage III.1: Encoder warm-up with a fixed decoder: Encoder warm-up first freezes the one-step decoder, backpropagating the complete reconstruction and perceptual objective through it while optimizing only the encoder.This aligns the encoder with the distilled decoder without perturbing the decoder’s reconstruction ability.
  • Stage III.2: End-to-end fine-tuning: End-to-end fine-tuning then unfreezes the decoder, enabling encoder–decoder co-adaptation while the KL term maintains compatibility with the anchor-latent space.The resulting Mage-VAE provides one-step encoding and decoding, high-fidelity reconstruction, and a compact, regularized latent space for subsequent Mage-Flow training.

B VAE Latency Across Resolutions

Mage-VAE is the fastest evaluated tokenizer across resolutions, with its latency advantage widening sharply at high resolution. Its single-step, memory-efficient design supports native-resolution packing and practical high-resolution generation under fixed compute budgets.

  • Cross-resolution comparison: Mage-VAE is the fastest at every tested resolution, and its advantage widens sharply as resolution grows.At 512^2, methods are broadly comparable, but the gap expands substantially at larger resolutions.
  • High-resolution latency: 149.6/375.0 ms: Mage-VAE encodes / decodes at 4096^2, while FLUX.2-VAE requires 47.3/183.9 s and Qwen-Image-VAE requires 106.7/192.0 s.At 2048^2, the latency gap reaches roughly an order of magnitude; other VAEs can also slow by orders of magnitude or run out of memory.
  • Latency evaluation: Table 17 compares VAE encoding and decoding latency from 512^2 to 4096^2 on an 80GB NVIDIA A100 GPU using bf16 precision.Each latency cell reports encoding / decoding time in milliseconds, with lower values better.
  • System implications: Mage-VAE’s single-step, memory-efficient design makes native-resolution packing and high-resolution generation practical under a fixed compute budget.The tokenizer must repeatedly encode and decode large images during native-resolution training and high-resolution inference, directly affecting end-to-end throughput.

C Recaptioning System Prompt

The recaptioning system uses one Qwen3-VL-32B-Instruct prompt to produce four granular caption layers in a single JSON response. It enforces objective, physically consistent descriptions through a defined procedure, schema, and style guide.

  • Prompt Design: Qwen3-VL-32B-Instruct generates entity, phrase, composition, and photographic descriptions in one pass as a JSON object.The system prompt produces the four caption layers at multiple granularities for each retained image.
  • Prompt Design: The prompt requires absolute objectivity, physical and logical consistency, and structured descriptions grounded in visible content.It prohibits subjective wording and requires real-world spatial and physical consistency, such as shadows aligning with light sources.
  • Captioning Procedure: The captioning procedure identifies the main subject, describes the background, details composition and style, extracts subject phrases, and summarizes the image.The steps specify subject attributes including color, quantity, and position, along with environment, lighting, atmosphere, medium, and perspective.
  • Output Schema: The JSON schema defines Entity, Phrase, Composition, and Photographic fields with distinct requirements for identification, keywords, context, and technical description.The Photographic field includes image category, inferred style, OCR text, entities, lighting, color, camera composition, and photography techniques.
  • Style Guide: The style guide infers a visual style, defaults to photorealistic photography when unclear, and applies the style consistently across Composition and Photographic.The Photographic description must mention the style explicitly and reinforce it through lighting, perspective, and atmosphere details.

D Reward Models

Diffusion-NFT uses four capability-routed reward evaluators that convert generation or editing outputs into scalar rewards in [0, 1]. Three evaluators score text rendering, aesthetic quality, and semantic understanding, while RationalRewards scores instruction-based editing.

  • Text rendering: Text-rendering reward averages normalized target-string scores based on PaddleOCR-VL-1.5 recognition and capped character-level Levenshtein distance.A perfect rendering scores 1, while an absent or unrecognizable target scores 0.
  • Aesthetic quality: Aesthetic reward averages independent binary Qwen3.5-27B judgments across photographic-quality and human-anatomy criteria.The reward takes evenly spaced values in [0, 1], and non-applicable criteria return 1 by convention.
  • Semantic understanding: Semantic reward uses Qwen3.5-27B to average yes/no alignment checks covering objects, attributes, counts, spatial relations, and actions or interactions.Fine-grained faithfulness checks provide a dense, interpretable signal for multi-object scenes.
  • Instruction-based editing: RationalRewards rates instruction-based edits on four 1–4 aspects: text faithfulness, image faithfulness, physical and visual quality, and text rendering.The normalized score gives an all-4 edit a reward of 1 and an all-1 edit a reward of 0.

E Application: Scientific Diagram Generation

Mage-Flow-Base transfers efficiently to scientific diagram generation, where joint fine-tuning improves structural accuracy and enables a 4B model to produce diagrams comparable to the 9B SciForma baseline. The resulting model handles dense layouts, directional arrows, and legible text across diverse scientific schematic types.

  • Motivation: Scientific diagram generation stress-tests models on synchronized text rendering, directional arrow routing, and global layout correctness.The task is presented as a representative, structure-critical domain for evaluating transferability.
  • Joint Training Strategy: A single-stage balanced mixture combines SciForma-700K with natural scenes, targeted text-rendering subsets, and poster designs.The recipe aims to preserve general generative fidelity while building architectural competence, unlike SciForma’s original two-stage fine-tuning.
  • Quantitative Results: 61.61 Overall score: Mage-Flow-SciForma outperforms zero-shot Mage-Flow-Base at 40.80 by 20.81 points on SciFormaBench-2K.Component and arrow accuracies also increase by 13.70 and 24.70 points, respectively.
  • Qualitative Analysis: Mage-Flow-SciForma generates neural-network skip-connections, unrolled transformer blocks, directional processing pipelines, and legible mathematical indices.Figure 26 also shows text-annotated workflows and step-by-step descriptors inside nodes and dense workflows.
  • Qualitative Analysis: After joint fine-tuning, Mage-Flow-SciForma corrects layout issues and produces precise text and directional arrows, matching the 9B SciForma-9B-Base vertical baseline.Zero-shot Mage-Flow-Base has basic layout components but weaker text legibility and arrow semantics.
Loading 2607.19064v2…