Ovis-Image Technical Report

TL;DR

Ovis-Image is Alibaba AIDC-AI’s open 7B text-to-image diffusion model (Apache-2.0, Nov 2025) built on top of Ovis-U1, specifically optimized for high-quality text rendering under tight compute budgets. The headline claim is text-rendering quality competitive with Qwen-Image (20B+7B) and GPT-4o on CVTG-2K and LongText-Bench while running on a single high-end GPU. Paired with a 2B Ovis2.5 understanding module that acts as the text conditioner (effective 2B+7B configuration), it ships an OvisImagePipeline in diffusers, weights on Hugging Face, and a citation pointing at the technical report (arXiv:2511.22982).

Key claims

Ovis-Image is a 7B text-to-image model “specifically optimized for high-quality text rendering, designed to operate efficiently under stringent computational constraints,” built upon Ovis-U1 [HF model card §Overview].
At a 2B+7B effective configuration, Ovis-Image reaches CVTG-2K Word Accuracy averages of 0.9200 (vs Qwen-Image 0.8288 at 7B+20B, GPT-4o 0.8569, Seedream 3.0 0.5924, SD3.5 Large 0.6548), with NED 0.9695 and CLIPScore 0.8368 [HF model card §Performance, CVTG-2K table].
On LongText-Bench, Ovis-Image scores 0.922 EN and 0.964 ZN — comparable to Qwen-Image (0.943 / 0.946) and surpassing GPT-4o (0.956 / 0.619) on the Chinese split, while large open models like FLUX.1-dev (0.607 / 0.005), HiDream-I1-Full (0.543 / 0.024), and BAGEL (0.373 / 0.310) fall sharply on logographic text [HF model card §Performance, LongText-Bench table].
On the general DPG-Bench composition score, Ovis-Image hits 86.59 overall (Entity 92.38, Relation 93.98), ranking above Ovis-U1 (83.72), OmniGen2 (83.57), HiDream-I1-Full (85.89) and just behind Qwen-Image (88.32) and Seedream 3.0 (88.27) despite the much smaller parameter budget [HF model card §Performance, DPG-Bench table].
On GenEval, Ovis-Image’s overall score is 0.84 — at parity with GPT-4o and Seedream 3.0, and within 0.03 of Qwen-Image 0.87 / Ovis-U1 0.89; standout categories are Single Object (1.00) and Attribute Binding (0.80) [HF model card §Performance, GenEval table].
On OneIG-Bench-EN the Text-rendering sub-score is 0.914 — the highest in the comparison set, including Qwen-Image (0.891), GPT-4o (0.857), HunyuanImage-2.1 (0.816), and Seedream 3.0 (0.865) — driving a 0.530 overall tied with Seedream 3.0 and within 0.009 of Qwen-Image [HF model card §Performance, OneIG-EN table].
On OneIG-Bench-ZN the Text sub-score is 0.961 (vs Qwen-Image 0.963, GPT-4o 0.650, HunyuanImage-2.1 0.896), confirming text-rendering parity with Qwen-Image on the Chinese side at roughly one-third the active parameter count [HF model card §Performance, OneIG-ZN table].
The architecture is two-part: an Ovis2.5-2B understanding/conditioning module plus a 7B diffusion generator, packaged in diffusers as OvisImagePipeline with bfloat16 inference, 50 denoising steps, and CFG scale 5.0 as the default recipe [HF model card §Inference with Diffusers].
The model is licensed under Apache-2.0, ships safetensors weights (ovis_image.safetensors, ae.safetensors, Ovis2.5-2B) on Hugging Face, and is supported by Draw Things and DiffusionBee for local inference at release [HF model card §License, §Local Apps].
The code base is built on Ovis (AIDC-AI/Ovis) and FLUX (black-forest-labs/flux) according to the acknowledgements [HF model card §Acknowledgments].

Method

Ovis-Image follows the unified-multimodal lineage from Ovis-U1 Technical Report but specializes the generation side, dropping the understanding and editing heads. The HF model card describes a 2B Ovis2.5 understanding module — derived from the Ovis MLLM series — feeding conditioning into a 7B FLUX-derived diffusion backbone with a separately stored VAE (ae.safetensors). The pipeline ships a single OvisImagePipeline class in a fork of the HuggingFace diffusers library, with the standard from_pretrained → pipe(prompt, ...) interface; the reference invocation uses bfloat16 weights, 50 denoising steps, and CFG scale 5.0. Inference fits on a single high-end GPU.

The benchmark methodology compares Ovis-Image against the current frontier of open and closed image generators on five evaluation axes: CVTG-2K and LongText-Bench for typography rendering, DPG-Bench and GenEval for compositional alignment, and OneIG-Bench (EN + ZN) for an aggregate alignment / text / reasoning / style / diversity score. The reported parameter counts use the “X+Y” convention (understanding module + generator), allowing direct comparison against Qwen-Image’s 7B+20B, OmniGen2’s 3B+4B, BAGEL’s 7B+7B, and HiDream-I1-Full’s 11B+17B configurations.

The architectural details beyond the diffusers pipeline (specifically how the 2B understanding module is connected to the 7B generator, whether it follows Ovis-U1’s bidirectional token refiner or a different conditioning route, and what training curriculum produces the text-rendering specialization) are documented in the arXiv tech report (2511.22982), which was unreachable at filing time; this page summarizes only the HF model card claims.

Results

The headline result is text-rendering parity with — and on several splits, dominance over — much larger systems. On CVTG-2K average Word Accuracy across 2/3/4/5-region prompts, Ovis-Image scores 0.9200 vs Qwen-Image’s 0.8288 (+9.12 pp gap in Ovis-Image’s favor) and GPT-4o’s 0.8569 (+6.31 pp), at roughly 30% of Qwen-Image’s combined parameter count. The CVTG-2K NED of 0.9695 is the highest in the comparison table.

On LongText-Bench, Ovis-Image’s 0.964 ZN score exceeds Qwen-Image (0.946) and GPT-4o (0.619), reinforcing the Chinese-text-rendering specialization that Qwen-Image’s recipe also exhibits. The general-composition scores (DPG-Bench 86.59, GenEval 0.84) place it within striking distance of the 20B-class frontier while remaining clearly above the previous-generation 7B class (OmniGen2, Janus-Pro, Ovis-U1).

The OneIG-Bench tables also encode a recurring pattern: text-rendering scores for the open frontier are bimodal — models with explicit text-rendering curricula (Qwen-Image, Ovis-Image, HunyuanImage-2.1) cluster at 0.8+ while equally-large general-purpose models (BAGEL 0.244, Janus-Pro 0.001 EN, BLIP3-o 0.013 EN) fall to near-zero. Ovis-Image’s contribution is hitting the high cluster at the smallest active parameter budget on the table.

Why it’s interesting

Ovis-Image is the smaller, text-rendering-specialized successor to Ovis-U1 Technical Report from the same AIDC-AI team, and the direct open competitor to Qwen-Image Technical Report on the typography axis — at roughly 30% of Qwen-Image’s combined parameter count it reports CVTG-2K and LongText-Bench numbers in the same range, suggesting the Qwen-Image text-rendering recipe (synthetic typographic data + curriculum + decoder fine-tune on text-rich corpora) generalizes to smaller backbones when paired with a sufficiently capable VLM-as-conditioner (Ovis2.5-2B here vs Qwen2.5-VL-7B there). For the Open foundation-model releases cluster, it’s a new datapoint in the “single-model release with deep variant packaging” pattern: Apache-2.0 weights, diffusers integration, local-app integrations (Draw Things / DiffusionBee), HF demo Space, and a citation-ready arXiv report — all coordinated at release. The interesting open question for the wiki is whether the 2B+7B configuration becomes the new operating point for layout-sensitive applications (posters, UI mockups, infographics) where the 20B+ models have until now been the only credible open option.