Ovis-Image Technical Report

• Ovis-Image addresses a persistent challenge: reliable, high-quality in-image text rendering at low computational cost requires both fine-grained visual synthesis and robust language comprehension.Strong text rendering has typically been associated with very large models or closed-source systems.Source: §1 · p. 2
• Unlike prior 3B Ovis-U1, which struggled with artifacts, hallucinations, and in-image text quality, Ovis-Image prioritizes text rendering while preserving general visual fidelity at acceptable cost.The design responds to systems that combine strong multimodal backbones with tens of billions of parameters, increasing training and deployment costs.Source: §1 · p. 2
• At 7B parameters, Ovis-Image matches 20B-class Qwen-Image and approaches GPT4o and Gemini on text-centric tasks.It upgrades the multimodal backbone to Ovis 2.5 and couples it with a diffusion-based visual decoder.Source: §1 · p. 2
• Ovis-Image produces legible, correctly spelled, semantically consistent text on layout-sensitive prompts without compromising overall visual quality.The report highlights posters, banners, logos, UI mockups, and infographics across diverse fonts, sizes, and aspect ratios.Source: §1 · p. 2
• Its 7B parameter budget and streamlined architecture support single-high-end-GPU deployment, low-latency interactive use, and batch production serving.This is intended to make near-frontier text rendering practical where models with tens of billions of parameters are impractical.Source: §1 · p. 2
• The report argues that appropriate architecture and text-centric training on a strong multimodal backbone can approach frontier-level performance within a compact 7B footprint.It concludes that high-quality in-image text rendering does not inherently require extremely large models.Source: §1 · p. 2

Architecture

• Ovis-Image builds on Ovis-U1, streamlining its structure while increasing MMDiT parameters (Wang et al., 2025).Figure 2 presents the overall architecture; Table 1 summarizes the modules.Source: §2 · p. 2, Figure 2 · p. 3

MMDiT & VAE

• The visual decoder uses MMDiT with RoPE and flow matching, comprising 6 double-stream and 27 single-stream blocks with 24 attention heads.It uses SwiGLU, and the FLUX.1-schnell VAE remains frozen during training (Esser et al., 2024; Su et al., 2024; Verd´u & Mart´ın, 2024; Dauphin et al., 2017; Labs, 2024).Source: §2 · p. 3

Text Encoder

• Ovis2.5-2B supplies text conditioning through its final hidden states, outperforming Qwen2.5-VL-7B on OpenCompass; Ovis-Image removes Ovis-U1’s refiner (Lu et al., 2024; Lu et al., 2025; Bai et al., 2025).Ovis is trained on multimodal datasets to align visual and textual representations.Source: §2 · p. 3

3.1 Data for Pre-training

• Pre-training combines heterogeneous image–text sources with bilingual recaptioning, quality filtering, deduplication, and synthetic typography examples.The corpus covers photographs, illustrations, design assets, and UI-like mockups, including posters, banners, logos, and UI layouts. Synthetic rendering-engine samples provide controlled examples of fonts, sizes, and placements; filtering removes corrupted images, mismatched or uninformative captions, and policy-violating content.Source: §3.1 · pp. 3–4

3.2 Data for Supervised Fine-tuning

• Supervised fine-tuning uses a cleaner, typically 1024-pixel subset spanning aspect ratios, with synthetic examples and content balancing to support high-fidelity instruction-following generations.The subset emphasizes clean visuals and well-formed prompts, and balances content type, resolution, and aspect ratio to avoid overfitting to dominant patterns.Source: §3.2 · p. 4

3.3 Data for DPO

• DPO preferences pair a high-quality pool image with an SFT-generated candidate, then use combined scorer rankings to designate winners and losers.About 90% of the pool consists of pre-filtered generations of common objects and everyday scenes; the remaining ∼10% is an in-house collection of design and creative content. The automatic scorers include HPSv3 Ma et al. (2025), CLIP Radford et al. (2021), and PickScore Kirstain et al. (2023).Source: §3.3 · p. 4

3.4 Data for GRPO

• GRPO uses a compact, deliberately distinct prompt set to target accurate, legible Chinese and English text generation across varied layouts and difficulty levels.Prompts include posters, title cards, UI elements, and product labels, ranging from short slogans to longer multi-line phrases.Source: §3.4 · p. 4

4.1 Training Infrastructure

• The four-stage pipeline initializes each stage from its predecessor, while PyTorch HSDP, BF16 mixed precision, checkpointing and offloading, Flash Attention, and regional compilation support training efficiency.FP32 master weights are retained, and distributed checkpointing reduces model-state saving overhead.Source: §4.2 · p. 4, §4.1 · p. 4

4.2 Training Procedure

• Pretraining optimizes a randomly initialized MMDiT with frozen pretrained text encoder and VAE, using noise-prediction loss and progressing from 256 × 256 images to 512–1024-pixel images with aspect ratios of 0.25–4.0.AdamW, a constant learning-rate schedule, and brief linear warmup are used.Source: §4.2 · pp. 4–5
• Supervised fine-tuning adapts the pretrained model to instruction-style text-to-image data from open and proprietary sources, improving prompt faithfulness and aesthetics while training on varied aspect ratios up to 1024 resolution.It retains the noise-prediction objective and uses a small learning rate and shorter schedule to preserve pretrained visual competence.Source: §4.2 · p. 5
• DPO uses human and model preference pairs with a frozen SFT reference; winner-preserving Diffusion-SDPO (Fu et al., 2025) downweights conflicting loser updates, avoiding noisy artifacts and improving visual quality, prompt adherence, and text rendering.DPO also improves helpfulness and harmlessness.Source: §4.2 · p. 5
• GRPO (Shao et al., 2024; Liu et al., 2025) scores groups of sampled images with reward models, then runs approximately 500 steps using fewer denoising steps, coefficients-preserving sampling (Wang & Yu, 2025), and a KL constraint.The policy is optimized to maximize expected reward while limiting divergence from the DPO model.Source: §4.2 · pp. 5–6

Evaluation overview

• Across text-rendering and general-generation benchmarks, Ovis-Image consistently outperforms significantly larger open-source baselines while remaining parameter-efficient.The evaluation covers LongText-Bench, CVTG-2K, DPG-Bench, GenEval, and OneIG-Bench.Source: §5 · p. 6

CVTG-2K

• Ovis-Image achieves the highest overall word accuracy across CVTG-2K regions, as well as the highest NED and CLIPScore.CVTG-2K uses 2,000 prompts requiring 2 to 5 English text regions per image.Source: Table 2 · p. 6, §5 · p. 6

LongText-Bench

• On LongText-Bench, Ovis-Image performs best on Chinese text and renders long English text comparably to closed-source and larger models.The benchmark evaluates long-text rendering in both English and Chinese.Source: §5 · p. 6, Table 3 · p. 7

DPG-Bench

• On DPG-Bench’s 1,000 dense prompts, Ovis-Image delivers robust performance against closed-source models and larger open-source models.The benchmark examines prompt adherence across multiple dimensions.Source: §5 · p. 6, Table 4 · p. 7

GenEval

• GenEval results show that Ovis-Image has competitive controllable generation capabilities on compositional, object-centric prompts.GenEval covers prompts with a wide range of object attributes.Source: Table 5 · p. 7, §5 · p. 7

OneIG-Bench

• On OneIG-Bench, Ovis-Image demonstrates exceptional bilingual performance, particularly in the text dimensions.The evaluation spans multiple dimensions of text-to-image generation.Source: §5 · p. 7, Table 6 · p. 8, Table 7 · p. 8

Computational Overhead

• Ovis-Image offers a favorable trade-off between inference time and GPU memory, with substantially lower memory use and faster inference than larger baselines.The comparison uses 1024×1024 images, 50-step sampling, and BF16 inference.Source: §5 · p. 7, Table 8 · p. 8

6 Conclusion

• Ovis-Image achieves text rendering quality comparable to much larger open models and approaches leading closed-source systems while preserving solid general-purpose generation.The report attributes this result to organizing architecture, data curation, and alignment objectives around in-image text rather than treating text as a byproduct of generic image synthesis.Source: §6 · p. 8
• Its diffusion-based visual decoder, Ovis 2.5 multimodal backbone, and text-centric training pipeline enable strong text rendering at 7B scale.The model fits on a single high-end GPU.Source: §6 · p. 8