MOSS-Image Technical Report
MOSI. AI
cs.CV
2026-10-16
A 5B DiT pretrained with layout-aware text supervision plus a GPT-6 planning agent hits 96.3% coverage on page-length bilingual text, first among open-weight models.
Closed-source image generators now handle information-dense graphics: infographics, UI mockups, knowledge cards. Open-weight models still fail on two fronts. Long text comes out garbled, and layout elements land in the wrong places. The report splits this gap into two coupled problems. Rendering means faithfully drawing long text and spatial constraints. Specification means expanding a one-line request into detailed content and layout instructions. Neither alone is enough: a detailed spec is useless if the model cannot draw it, and strong rendering cannot decide the content and design choices the user left unspecified. The system is built directly on that diagnosis.
The backbone is a 5B single-stream DiT pretrained from scratch. Text tokens come from a frozen Qwen3.5-9B encoder, image latents from a frozen FLUX.2 VAE; RoPE is extended to four axes, one reserved for future reference-conditioned generation. The load-bearing design decision: natural-language descriptions and structured layout JSON share a single text-conditioning pathway, so the model treats both input forms the same way and the agent's layout output needs no separate interface.
Three parts of the training recipe stand out:
At inference, an agentic prompt enhancer built on GPT-6 routes requests by medium (poster, infographic, UI), retrieves knowledge and visual references, and emits a JSON spec with literal text, elements, bounding boxes, and typography, validated before it enters the same conditioning pathway. Unlike generator-agnostic agents that can only emit prose prompts, it issues structured layouts directly.
| Benchmark | Baseline | Result |
| LLTB Mean-6 (in-house, mean 1,340-char targets) | Qwen Image 2512: 64.9; Ideogram 4: 63.5 | 96.3, best open-weight |
| LLTB longest bin | EN 97.1, ZH 92.4 | |
| 7Bench layout mIoUdet / AUC | Ideogram 4: 0.44 / 0.20; GPT Image 2: 0.30 / 0.16 | 0.88 / 0.70, best overall |
| RLA end-to-end Elo (61 tasks, blind human votes) | GPT Image 2: 1844.4; Seedream 5.0 Pro: 1661.2 | 1698.7, second |
| Backbone Elo under shared enhanced prompts | ERNIE Image: 1748.5 | 1807.3, first |
The prompt-strategy ablation separates the two layers: short requests get 1081.7 Elo, detailed prompts 1509.2, and the PE's JSON layout spec 1622.9. Rewriting the same spec as prose drops it to 1506.2, so structure buys about 117 Elo; under shared enhanced natural-language prompts the backbone climbs further to 1807.3. The paper also declares LongTextBench saturated, with strong models bunched between 96.05 and 98.85 on targets averaging 120 characters. LLTB's targets average 1,340 characters, 11 times longer.
This pushes open-weight results close to the closed-source frontier on page-length text and layout control, and the recipe is disclosed in detail: data retention, line-break supervision, and coordinate representation each carry ablations. The finding that zero-centered absolute bounding boxes generalize to unseen resolutions better than [0, 1000] normalized coordinates is directly reusable. Teams building poster, infographic, or UI generation can borrow the backbone-plus-planning-agent architecture and the JSON layout interface wholesale. The honest framing: individual pieces are incremental, and the report reads more like a well-executed integration study than a conceptual breakthrough. Its value lies in the completeness of the recipe and the released benchmarks.
The authors list two: the backbone supports neither reference-image conditioning nor image editing, and the prompt enhancer depends on an external language model whose aesthetic judgment and factual errors propagate into the spec. That enhancer is closed-source GPT-6, so the inference chain of this open system is not fully open. On the evaluation side: RLA has only 61 tasks, and MOSS leads Seedream 5.0 Pro by roughly 37 Elo while its own confidence interval spans about 85 points, so second place is not statistically settled. LLTB coverage is word-multiset recall, which ignores order and does not penalize extra characters, so a high score does not guarantee correct layout. Both LLTB and RLA are self-built and self-scored, without community replication so far, and the cost and latency of GPT-6 planning are not reported. The paper also states plainly that backbone results characterize the model as a whole without isolating the causal contribution of individual training choices.