The Pulse
MOSS-Image Scores 96.3% on Long-Text Rendering Benchmark
OpenMOSS released MOSS-Image, a 5-billion-parameter image model built to render long English and Chinese text and follow layout instructions. Its technical report reports 96.3% average text coverage on a new benchmark, while noting limits i
AI.info Team · Written with AI · fact-checked ·
MOSS-Image recorded 96.3% average text coverage on a new benchmark designed to test paragraphs, not just a few words on a poster. OpenMOSS says the 5-billion-parameter model leads the open-weight systems it evaluated on long English and Chinese text rendering. The result comes with an important qualification: the benchmark measures how much target text an optical character recognition system can recover, not whether every word appears in the right order.
OpenMOSS released the model weights and code on October 16, 2026. Its technical report, submitted to arXiv that day, describes the model and the evaluations behind the headline score. The project’s GitHub page marked the report as available on October 19, extending the initial release with a detailed account of the system’s design and tests.
What the 96.3% score measures
The team built LongLongTextBench, or LLTB, with 240 image-generation tasks split evenly between English and Chinese. Each task asks a model to render a literary passage as text in an image; the full targets average about 1,340 characters. The researchers generated four images for each task and included all four in scoring, rather than selecting the best result.
The report says, “MOSS-Image achieves state-of-the-art performance among open-weight models on LLTB, with an average text coverage of 96.3%.” The authors of “MOSS-Image Technical Report” compare that result with 64.9% for Qwen Image 2512 on the same benchmark. MOSS-Image’s score on the longest Chinese passages is lower, at 92.4%, compared with 97.1% on the longest English passages.
Coverage is not exact transcription accuracy. The evaluation uses OCR to match recognized English words and Chinese characters against target text, but it does not check word order or directly penalize extra recognized text. That makes 96.3% a measure of recovered text units under the benchmark’s scoring rules—not a claim that nearly every generated image contains a perfectly correct passage.
JSON layouts give the model coordinates
MOSS-Image accepts ordinary prompts as well as structured JSON describing objects, text and bounding boxes. Users can specify both what should appear and where it belongs, giving the image generator spatial instructions alongside its text prompt. The model’s base and fine-tuned versions share the same architecture and sampling interface; OpenMOSS says the SFT version produces clearer text grouping in its example comparisons.
The report also tests object placement using 7Bench. Among detected objects, MOSS-Image posts a mean intersection-over-union score of 0.88 and a layout-area-under-curve score of 0.70, above the corresponding reported scores for Ideogram 4 in the team’s evaluation. Those measurements are conditional on detections: when the researchers include every target object, including objects the system fails to detect, MOSS-Image’s mean IoU is 0.68. The distinction matters because a high placement score for detected objects does not mean the model places every requested element.
A strong system score relies on an external planner
OpenMOSS also tests complete systems on RealLayoutArena, a set of 61 short design requests covering posters, infographics and user interfaces. In blind pairwise comparisons, raters judge instruction following, rendering quality and spatial organization. MOSS-Image ranks second in that end-to-end comparison, behind GPT Image 2.
The full system pairs the image model with an intent-aware prompt enhancer that turns a brief request into detailed content and layout instructions. The report says that planner uses GPT-6 at inference time. In a separate comparison where several image generators receive shared enhanced prompts, the MOSS-Image backbone ranks first. Those results point to two distinct contributions—the open image generator and a planning stage that depends on an external model—rather than a wholly self-contained open system.
Open weights, with editing still absent
The released weights use the Apache-2.0 license, and the repository includes a Python inference path and a ComfyUI plugin. The model is built around a single-stream diffusion transformer, with a frozen Qwen3.5-9B text encoder and a FLUX.2 VAE, according to the report. OpenMOSS lists support for prompts up to 4,096 tokens and recommends an NVIDIA GPU for its documented local setup.
The team identifies a clear boundary: the current backbone generates images from text and layouts but does not directly accept reference images or support image editing. Its planner can inspect visual references to help prepare a prompt, yet those images do not condition the generator itself. MOSS-Image’s headline benchmark result therefore speaks to a focused capability—rendering dense text and arranging elements—not a general ability to revise an existing image.