- The paper introduces UNO, a lightweight post-training method that uses frozen language-captioning and visual-feature objectives to steer noisy generative representations toward stronger semantic alignment.
- UNO improves BAGEL-7B by 3.4 points on GenEval2, 3.5 points on UniGenBench++, and up to 0.70 points on GEdit-Bench while preserving understanding performance.
- The method generalizes to Show-o, complements self-consistency chain-of-thought prompting, and requires only 5,000 training iterations, though it does not by itself improve world-knowledge retrieval.
Motivation and problem statement
Unified Multimodal Models (UMMs) integrate language comprehension, visual understanding, and visual generation within a single framework, but state-of-the-art systems such as BAGEL increasingly adopt decoupled representations—separate vision encoders, FFNs, or disjoint transformer experts—to avoid optimization conflicts between high-level semantic abstraction and low-level generative objectives. While effective for task-specific performance, this decoupling induces a strictly unidirectional information flow: the generation expert is conditioned on the understanding expert's outputs but receives gradients only from flow-matching losses. The authors' central observation is that this leaves the understanding expert's strong semantic capabilities under-exploited as a supervisory signal. They hypothesize that explicitly supervising generative representations through understanding objectives operated by the model's own frozen understanding expert can inject semantic constraints that improve generation.
The UNO framework
Understanding-Oriented Post-Training (UNO) is a lightweight post-training stage applied to a pretrained UMM (BAGEL-7B) with the understanding expert frozen. Two complementary proxy objectives supervise the noised intermediate generative representations:
- Language supervision: the understanding expert is conditioned on noised generation representations rather than native vision tokens and is trained to caption the image via next-token prediction. Because the model can trivially "copy" prompt tokens stored in high-capacity visual representations, the authors mask conditional prompt tokens from the supervision stream and apply semantic augmentation—re-captioning target images with an alternative captioner (Qwen2.5-VL-7B) so supervision text is semantically consistent but lexically different from the conditioning prompt.
- Visual understanding supervision: learnable metaquery tokens (256 per sample, following MetaQuery) are processed by the frozen understanding expert, and their output hidden states are regressed onto dense SigLIP features of the target image via cosine similarity.
The total objective combines these with the standard flow-matching MSE loss (λ1=0.1, λ2=0.2), using packed sequences with modified attention masks to prevent information leakage. Training is short—5K iterations on roughly 2M curated text-image pairs for generation (excluding distillation-based data such as BLIP3o-60k to avoid benchmark template leakage) and CrispEdit-2M for editing—with token overhead limited to ~336 additional tokens per sample.
Main results
UNO yields consistent gains over BAGEL across both tasks without degrading understanding performance:
| Benchmark |
BAGEL |
BAGEL + UNO |
Δ |
| GenEval2 |
71.7 |
75.1 |
+3.4 |
| DPG-Bench |
84.03 |
86.12 |
+2.09 |
| UniGenBench++ (overall) |
61.53 |
65.03 |
+3.50 |
| GEdit-Bench-EN (overall) |
6.52 |
7.17 |
+0.65 |
| GEdit-Bench-CN (overall) |
6.50 |
7.20 |
+0.70 |
On UniGenBench++, the largest per-dimension improvements appear precisely where semantic comprehension matters most: compound (+6.70), action (+5.51), attribute (+5.12), and relationship (+4.56). World knowledge slightly decreases (−1.42), which the authors attribute to training-data coverage rather than the method itself—a concession supported by their WISE evaluation, where UNO alone does not help (0.52 → 0.51) but adding 50K distilled knowledge samples recovers and exceeds the baseline (0.56). On editing, UNO achieves the best overall score on GEdit-Bench-EN among all compared models, including FLUX-Kontext and Step1X-Edit, and generalizes to Chinese instructions despite English-only training data.
Analysis
Comparison against post-training baselines. Under identical data settings, joint understanding supervision outperforms supervised fine-tuning (GenEval2 73.6 vs. 75.1; UniGenBench++ 63.02 vs. 65.03) and reconstruction alignment (RecA). Language supervision alone contributes more than visual supervision alone, and their combination is best, confirming complementarity between abstract linguistic signals and dense spatially structured ones.
Feature-space evidence. PCA visualization of generation-expert latents at heavily noised timesteps shows that UNO-trained representations exhibit less noise, better semantic organization, and preserved fine-grained details relative to flow-matching-only training. Per-layer gradient analysis shows understanding and generation gradients are largely orthogonal, with some positive alignment—i.e., the added objectives do not introduce optimization conflict.
Ablations. Several design choices are shown to matter: masking conditional prompts is essential (removing it drops GenEval2 from 75.1 to 73.4); more than 256 metaqueries degrades performance; unfreezing the understanding expert gives inconsistent gains and risks degrading its own benchmark performance; and the choice of vision connector (ViT similarity vs. MLP/DiT denoising targets) has little effect, so the simplest identity-projection cosine-similarity design is adopted.
Generality and interaction with CoT. Applied to Show-o, a discrete-diffusion unified model, UNO improves GenEval2 (58.57 → 61.26), DPG-Bench (80.04 → 81.94), and UniGenBench++ (50.39 → 53.41) while preserving MME understanding scores, indicating the framework is not specific to BAGEL's architecture. Notably, UNO outperforms self-CoT prompting alone (72.0 vs. 71.7 GenEval2 without CoT) and stacks with it (75.1 combined), suggesting representation-level and inference-time semantic enhancement are complementary mechanisms.
Limitations and open questions
The paper concedes several boundaries. UNO does not address vertical-domain capabilities such as world-knowledge retrieval, as evidenced by the flat WISE results without auxiliary data. The framework was validated only as a post-training stage on two base models at moderate scale; whether understanding supervision helps during pretraining, or scales differently with larger generators, remains untested. The mechanism is supported primarily by feature visualization and gradient-direction analysis rather than a causal account of why orthogonal gradients translate into benchmark gains. Finally, the slight regression on world-knowledge dimensions suggests the method may trade off coverage-dependent capabilities against semantic adherence, and the paper leaves open how to balance these without domain-specific data augmentation.
Conclusion
This work demonstrates that within decoupled UMMs, the understanding expert can serve as an effective supervisory signal for the generative pathway when information flow is re-routed so that understanding conditions on noised generative representations. With modest compute (5K iterations), no architectural changes, and a frozen understanding expert, UNO delivers consistent gains on generation and editing benchmarks while preserving understanding ability and generalizing across architectures. The result supports the position that capability synergy in unified models need not require fully shared representations, and identifies gradient-flow design—not parameter sharing—as a productive axis for deeper integration.