Uni-Layout: Unified Layout Generation & Evaluation

• Uni-Layout is a unified framework that treats layout synthesis as a conditional modeling problem across diverse design tasks.
• It integrates an autoregressive multimodal generator with a large-scale human feedback dataset and a learned evaluator to align with expert judgments.
• Dynamic-Margin Preference Optimization adapts the strength of layout preferences, enhancing design quality through targeted fine-tuning.

Uni-Layout denotes a research program in which layout generation is treated as a unified conditional modeling problem rather than a collection of task-specific systems. In its narrowest sense, the term refers to the framework introduced in “Uni-Layout: Integrating Human Feedback in Unified Layout Generation and Evaluation,” which combines a unified generator, a large human-feedback dataset, a human-mimicking evaluator, and a preference-optimization stage within one pipeline (Lu et al., 4 Aug 2025). In a broader sense, the term also aligns with a family of recent methods that unify layout generation across tasks, domains, or control interfaces by using shared representations, masked conditioning, diffusion denoisers, or autoregressive prompting (Zhang et al., 19 Feb 2025, Hui et al., 2023).

1. Scope and task taxonomy

The immediate motivation for Uni-Layout is twofold. First, many layout-generation systems are narrowly specialized: different architectures are commonly built for document layout generation, poster layout with a given background, product poster layout with constrained product elements, or poster design with both a background and specific text content. Second, the standard evaluation metrics used in layout generation are often described as perceptually misaligned with expert judgment; a layout may score better numerically while appearing worse to a human viewer (Lu et al., 4 Aug 2025).

To organize the task space, Uni-Layout introduces a two-dimensional taxonomy based on whether the background and element contents are free or constrained. The resulting four categories are presented as a unified task taxonomy covering representative layout-generation scenarios (Lu et al., 4 Aug 2025).

Category	Background	Element contents
BFEF	Free	Free
BCEF	Constrained	Free
BFEC	Free	Constrained
BCEC	Constrained	Constrained

Within this taxonomy, BFEF corresponds to background-free and element-free generation, BCEF to background-constrained and element-free generation, BFEC to background-free and element-constrained generation, and BCEC to simultaneous background and element-content constraints. Uni-Layout trains one generator across all four settings rather than introducing a different decoder or structural output space for each one (Lu et al., 4 Aug 2025).

The representative datasets used for these four task families are also explicit. BCEF is instantiated by CGL-Dataset, BCEC by CGL-Dataset V2, BFEF by PubLayNet, and BFEC by EP-Layout, which is introduced because prior BFEC datasets were not publicly available (Lu et al., 4 Aug 2025). This arrangement makes the claim of “unified” generation concrete: all task families are converted into a common layout-sequence formulation and addressed by one model.

2. Generator architecture and the broader unification paradigm

The Uni-Layout generator is built on LLaVA as a base MLLM rather than on diffusion. It is an instruction-following autoregressive multimodal model $G_{\Theta}$ trained with next-token prediction over layout sequences (Lu et al., 4 Aug 2025). Its inputs are visual inputs $\mathbf{I}_\mathrm{v}$, which may include a background image or element images depending on the task, and a textual task instruction $\mathbf{X}_\mathrm{instruct}$. Its output is a structured layout sequence $l$.

The prompt format is explicitly specified as

$$\mathrm{Human}:\mathbf{I}_\mathrm{v}<\backslash n\mathrm{>}~\mathbf{X}_\mathrm{instruct}\text{<STOP> Assistant}:l\mathrm{<STOP>}.$$

For BFEF tasks without image input, $\mathbf{I}_\mathrm{v}$ and the line-break token are omitted (Lu et al., 4 Aug 2025). The instruction itself is generated by a unified template

$$\mathbf{X}_\mathrm{instruct} = f_{\mathrm{instruct}}(T, b_a, b_c, e_a^1, \ldots, e_a^m, e_c^1, \ldots, e_c^n, O),$$

where $T$ is task description, $b_a$ background attributes, $b_c$ background content, $\mathbf{I}_\mathrm{v}$0 element attributes, $\mathbf{I}_\mathrm{v}$1 element contents, and $\mathbf{I}_\mathrm{v}$2 the output-format specification (Lu et al., 4 Aug 2025). The output sequence follows a structured layout format exemplified as

$\mathbf{I}_\mathrm{v}$3

This design belongs to a wider unification trend. LGGPT introduces Arbitrary Layout Instruction (ALI) and Universal Layout Response (ULR) as a uniform I/O template across multiple tasks and four layout domains, and uses Interval Quantization Encoding (IQE) to remove placeholders while preserving attribute identity (Zhang et al., 19 Feb 2025). LDGM instead defines unification through a diffusion view in which each layout attribute may be precise, coarse, or missing, so unconditional generation, completion, refinement, and mixed cases become instances of reversing a partially corrupted layout (Hui et al., 2023). LACE uses masked input and continuous diffusion to unify unconditional generation, conditional generation, completion, and refinement under one denoising model while adding differentiable alignment and overlap constraints (Chen et al., 2024). UniLayDiff treats partial constraints and relations as explicit modalities in a dual-path MM-DiT for content-aware layout generation (Liu et al., 9 Dec 2025). PlanGen unifies layout planning and image generation by representing layouts as token sequences with <grounding>, <ref>, and <box> tags inside one autoregressive transformer (He et al., 13 Mar 2025). OmniLayout-LLM pushes the same agenda into document AI with a coarse-to-fine curriculum over OmniLayout-1M, a million-scale multi-genre dataset (Kang et al., 30 Oct 2025).

This suggests that “unification” in recent layout research has become a technical design principle rather than a slogan. The recurring mechanisms are a common structured representation, a shared decoder or denoiser, and a conditioning interface that can absorb heterogeneous constraints without task-specific model branches.

3. Human feedback, Layout-HF100k, and the evaluator

A distinctive feature of Uni-Layout is that unification is not restricted to generation. The framework also introduces Layout-HF100k, described as the first large-scale human feedback dataset for layout generation, and a learned evaluator trained on that dataset (Lu et al., 4 Aug 2025).

Split	Total	Breakdown
Train	96,000	31,000 BCEF; 19,000 BFEF; 19,000 BFEC; 27,000 BCEC
Test	4,000	1,000 per task; approximately 1:1 positive/negative ratio

The dataset is built from model-generated candidate layouts that are labeled by professional annotators as qualified or unqualified. The annotation workflow has three stages: primary annotation, quality inspection by a separate verification team, and a statistical sampling audit of at least 10% of each batch by senior auditors. Annotators all have more than 5 years of graphic-design experience, and the workflow enforces a minimum label accuracy of 98% (Lu et al., 4 Aug 2025).

The evaluator $\mathbf{I}_\mathrm{v}$4 uses a dual-branch preprocessor $\mathbf{I}_\mathrm{v}$5 that converts a layout into both a visualized representation and an enriched geometric representation. The visualization branch renders different element types with distinct colored blocks, over the provided background image when present or over a blank canvas otherwise, and displays actual element content for BFEC and BCEC. The geometry branch extracts positions, sizes, spatial relations, and even color information from the visualization branch to form a structured prompt $\mathbf{I}_\mathrm{v}$6 (Lu et al., 4 Aug 2025).

The evaluator is also built on LLaVA. For quantitative prediction, it uses the hidden state $\mathbf{I}_\mathrm{v}$7 of the last token and a classifier head: $\mathbf{I}_\mathrm{v}$8 The positive-class probability $\mathbf{I}_\mathrm{v}$9 becomes the Layout Reward (LR) score (Lu et al., 4 Aug 2025). For qualitative reasoning, the evaluator produces chain-of-thought outputs in four stages: Layout Glimpse, Spatial Deconstruction, Aesthetic Appraisal, and Holistic Evaluation. CoT supervision is bootstrapped with GPT-4o for Stage 1 captions and DeepSeek-R1 for Stages 2 and 3, while human labels define the final conclusion (Lu et al., 4 Aug 2025).

The evaluator is trained with a combined objective

$\mathbf{X}_\mathrm{instruct}$0

On Layout-HF100k it reaches 85.5% accuracy, compared with 61.6% for GPT-4o, 57.8% for Claude3.5, and 54.3% for DeepSeek-R1. Its per-task accuracy is reported as 86.2% on BFEF, 87.2% on BCEF, 88.2% on BFEC, and 80.4% on BCEC (Lu et al., 4 Aug 2025).

4. Alignment through Dynamic-Margin Preference Optimization

Uni-Layout closes the loop between generation and evaluation with Dynamic-Margin Preference Optimization (DMPO). The generator produces two candidate layouts $\mathbf{X}_\mathrm{instruct}$1 and $\mathbf{X}_\mathrm{instruct}$2 for the same prompt; the evaluator selects a preferred layout $\mathbf{X}_\mathrm{instruct}$3 and a less preferred layout $\mathbf{X}_\mathrm{instruct}$4, and derives a confidence margin

$\mathbf{X}_\mathrm{instruct}$5

This preference strength is amplified by

$\mathbf{X}_\mathrm{instruct}$6

The resulting preference signal is then used in a DPO-style objective with an adaptive margin rather than a fixed one (Lu et al., 4 Aug 2025).

Conceptually, the method is designed to address a specific weakness of standard preference optimization: not all preferences are equally strong. A layout may be only slightly better than another, or dramatically better. DMPO uses the evaluator’s confidence gap to modulate how strongly the generator is pushed to separate them (Lu et al., 4 Aug 2025).

The training pipeline is staged. The unified generator is first pretrained by full-model fine-tuning for 10 epochs with a cosine learning-rate schedule and initial learning rate $\mathbf{X}_\mathrm{instruct}$7. The evaluator is then initialized from the pretrained generator weights and trained on Layout-HF100k with the same learning strategy. In the final stage, the evaluator is frozen and the generator is fine-tuned with LoRA for 3 epochs at learning rate $\mathbf{X}_\mathrm{instruct}$8. All experiments are run on a single node with 8 NVIDIA H100 GPUs (Lu et al., 4 Aug 2025).

The ablation evidence favors the adaptive-margin design. On Layout Reward, DMPO reaches 0.702, compared with 0.610 for DPO and 0.625, 0.667, 0.674, and 0.658 for fixed margins 0.5, 1, 1.5, and 2 respectively (Lu et al., 4 Aug 2025). The reported before/after alignment examples further indicate that DMPO reduces problematic overlaps and misalignments.

5. Empirical performance and relation to other unified systems

On the four task families, Uni-Layout reports the following task-specific metrics (Lu et al., 4 Aug 2025).

Task	Metrics	Uni-Layout
BFEF	Ove / Ali / Max.	0.001 / 0.00004 / 0.160
BFEC	Ove / Ali / Max.	0.00045 / 0.009 / 0.439
BCEF	$\mathbf{X}_\mathrm{instruct}$9 / $l$0 / $l$1	31.848 / 0.774 / 1
BCEC	$l$2 / $l$3 / $l$4	8.536 / 0.764 / 1

The human-centered results are central to the paper’s claims. Uni-Layout reports LR = 0.702, compared with 0.584 for GPT-4o, 0.575 for Claude3.5, 0.401 for DeepSeek-R1, and 0.422 for LLaVA. On Human Pass Rate (HPR), it reports 67.4%, compared with 56.9% for GPT-4o, 55.6% for Claude3.5, 37.7% for DeepSeek-R1, 40.3% for LLaVA, and 62.6% for the previous SOTA (Lu et al., 4 Aug 2025). The paper notes that HPR trends closely match LR, which is offered as evidence that the evaluator score is perceptually meaningful.

Within the wider unified-layout literature, several neighboring systems illustrate different technical emphases. LACE unifies unconditional generation, conditional generation, completion, and refinement in continuous diffusion, and reports PubLayNet refinement performance of FID 1.65 and MaxIoU 0.491 (Chen et al., 2024). LGGPT uses a compact 1.5B decoder-only LLM with ALI, ULR, and IQE, and is reported to beat prior 7B and 175B layout LLMs in the most extensive unified scenario (Zhang et al., 19 Feb 2025). UniLayDiff reports content-aware unconditional FID 3.15 on PKU and lower relation violation than RALF while supporting unconditional, conditional, completion, refinement, and relation tasks in one final model (Liu et al., 9 Dec 2025). OmniLayout-LLM, trained on OmniLayout-1M, reports Newspaper FID 39.73 in unconditional generation and 6.13 in $l$5, substantially surpassing prior document-layout experts on M$l$6Doc (Kang et al., 30 Oct 2025).

The empirical record therefore indicates that “unified layout” is not tied to a single architecture. It has been instantiated with MLLMs, decoder-only LLMs, continuous diffusion, discrete diffusion, and multimodal diffusion transformers.

6. Broader meanings, limitations, and adjacent traditions

The term “universal layout” also has older and more formal meanings. In rectangular layout theory, a rectangular layout is area-universal if and only if it is one-sided (0901.3924). The graph problem of deciding whether a plane graph admits such a layout is addressed by a polynomial-time algorithm in “A Polynomial Time Algorithm for Finding Area-Universal Rectangular Layouts” (Wang, 2013). For aspect ratios, the theory is sharper: a generic rectangular layout is weakly aspect ratio universal if and only if it is sliceable, and strongly aspect ratio universal if and only if it is one-sided and sliceable (Felsner et al., 2021). These results use “universal” in a combinatorial sense: a single layout topology can realize arbitrary area or aspect-ratio assignments while preserving equivalence.

A different systems tradition appears in GUI layout. “ORC Layout: Adaptive GUI Layout with OR-Constraints” states that ORC layout unifies grid layout and flow layout, supporting both their features as well as cases where grid and flow layouts individually fail (Jiang et al., 2019). Here, “unified layout” refers to a single responsive constraint language and solver rather than to generative modeling.

Recent multimodal work further broadens the term. PlanGen unifies layout planning, layout-to-image generation, image layout understanding, and layout-guided image manipulation inside one autoregressive vision-LLM (He et al., 13 Mar 2025). Uni-RS treats spatial layout as an explicit intermediate in a unified remote-sensing understanding-and-generation model, using Spatial-Layout Planning, Spatial-Aware Query Supervision, and Image-Caption Spatial Layout Variation to improve spatial faithfulness (Zhang et al., 25 Jan 2026).

This suggests that “Uni-Layout” now spans at least three research senses: formal universality of rectangular subdivisions, unified responsive specification in interface layout, and multi-task generative modeling of layouts in vision-language systems. The specific Uni-Layout framework of 2025 belongs to the third category. Its stated limitation is scope: it focuses on 2D graphic design layouts, and its proposed future direction is extension to 3D layout generation and evaluation, including new human-feedback mechanisms for VR, AR, and 3D modeling (Lu et al., 4 Aug 2025). More broadly, neighboring unified generators remain constrained by box-based representations, fixed or closed label sets, and domain-specific assumptions about overlap, background compatibility, or annotation schema (Chen et al., 2024, Kang et al., 30 Oct 2025).

In that broader research landscape, Uni-Layout is best understood as a human-centered consolidation of the unified-layout agenda: one generator, one evaluator, one feedback dataset, and one alignment mechanism, all organized around the claim that layout generation and layout evaluation should be learned together rather than separately (Lu et al., 4 Aug 2025).