---
title: Layout Grounding Score (LGS) Evaluation
url: https://www.emergentmind.com/topics/layout-grounding-score-lgs
type: topic
---

# Layout Grounding Score (LGS) Evaluation

Searching arXiv for the specified paper and closely related layout-to-image evaluation context.
arxiv_search.query({"search_query":"id:2509.16691 OR ti:\"InstanceAssemble: Layout-Aware Image Generation via Instance Assembling Attention\"","max_results":5,"sort_by":"relevance","sort_order":"descending"})
arxiv_search.query({"search_query":"ti:\"InstanceAssemble: Layout-Aware Image Generation via Instance Assembling Attention\"","max_results":10,"sort_by":"relevance","sort_order":"descending"})
arxiv_search.query({"search_query":"all:\"Layout Grounding Score\" OR all:\"DenseLayout\" OR all:\"layout-to-image benchmark\"","max_results":10,"sort_by":"relevance","sort_order":"descending"})
Layout Grounding Score (LGS) is an evaluation protocol for layout-to-image generation that decomposes grounding performance into one spatial component and three semantic components. Introduced with InstanceAssemble, LGS is designed to assess whether conditioned instances are both correctly localized and semantically faithful under dense layout conditions, using an open-set detector and a VQA model rather than a fixed closed-category taxonomy [2509.16691]. In its formal form, LGS is the 4-tuple $\bigl(\text{DetectIoU};\ \text{Accuracy}_{\mathrm{color}},\ \text{Accuracy}_{\mathrm{texture}},\ \text{Accuracy}_{\mathrm{shape}}\bigr)$, where spatial alignment is measured continuously over all conditioned instances and semantic correctness is evaluated only for instances whose predicted boxes satisfy an $\mathrm{IoU}>0.5$ gate.

## 1. Definition and conceptual scope

LGS is defined over a layout containing $N$ conditioned instances, each specified by a ground-truth box $b_i$ and a textual or visual description $c_i$ [2509.16691]. The generated image is evaluated instance by instance by querying an open-set detector, GroundingDINO, with the same instance description $c_i$ used in conditioning. For each instance, the detector returns a predicted box $b_i'$, and spatial agreement is measured by $\mathrm{IoU}(b_i', b_i)$.

The metric family separates spatial grounding from attribute grounding. The spatial term, DetectIoU, is the mean IoU over all conditioned instances. The semantic terms are per-attribute accuracies for color, texture, and shape, computed only on instances whose predicted boxes align sufficiently well. This design makes LGS a grounded evaluation rather than a purely appearance-based or purely localization-based score. A plausible implication is that the metric is intended to diagnose failure modes that conflate object presence, object placement, and object attributes.

## 2. Formal specification

Let $N$ denote the total number of conditioned instances. For each instance $i$, the detector output is $b_i'$, and the instance-level overlap is

$$
\mathrm{IoU}_i = \mathrm{IoU}(b_i',\, b_i).
$$

The spatial component is defined as

$$
\text{DetectIoU} = \frac{1}{N}\sum_{i=1}^{N}\mathrm{IoU}_i.
$$

Semantic evaluation is gated by localization quality. For each attribute $A \in \{\text{color}, \text{texture}, \text{shape}\}$, the predicted region $b_i'$ is cropped and passed to MiniCPM with a dedicated VQA query such as “What is the color of this object?” If the answer matches the condition $c_i$ on attribute $A$, the score is $s_i^A = 1$; otherwise $s_i^A = 0$ [2509.16691].

The semantic accuracy for attribute $A$ is then

$$
\text{Accuracy}_A
=
\frac{1}{\left|\{\,i:\mathrm{IoU}_i>0.5\}\right|}
\sum_{i:\mathrm{IoU}_i>0.5} s_i^A.
$$

Collecting the four components yields

$$
\mathrm{LGS}
=
\bigl(
\text{DetectIoU};\
\text{Accuracy}_{\mathrm{color}},\
\text{Accuracy}_{\mathrm{texture}},\
\text{Accuracy}_{\mathrm{shape}}
\bigr).
$$

This formalization gives LGS a mixed continuous-and-discrete structure: localization is averaged continuously through IoU, while semantic fidelity is averaged as binary correctness conditioned on sufficiently accurate localization.

## 3. Computation pipeline

The computation begins with detection. For each conditioned instance description $c_i$, GroundingDINO is run on the generated image using that same text prompt, and the top-scoring box $b_i'$ is retained if it passes the detector thresholds $\texttt{box\_thresh}=0.35$ and $\texttt{text\_thresh}=0.25$ [2509.16691]. If no box is returned, $\mathrm{IoU}_i$ is set to $0$.

Spatial accuracy is then computed directly from the returned boxes using

$$
\mathrm{IoU}_i=
\frac{\mathrm{area}(b_i' \cap b_i)}{\mathrm{area}(b_i' \cup b_i)},
$$

followed by averaging over all $N$ instances to obtain DetectIoU.

The next stage applies a semantic gate. The subset of well-localized instances is

$$
\mathcal{M} = \{\, i : \mathrm{IoU}_i > 0.5 \,\}.
$$

For each $i \in \mathcal{M}$ and each attribute in $\{\mathrm{color}, \mathrm{texture}, \mathrm{shape}\}$, the region inside $b_i'$ is cropped, a dedicated VQA query is asked, and the response is compared with the original condition $c_i$. Semantic scores are then averaged over $\mathcal{M}$, with no further weighting; every instance counts equally [2509.16691].

This stepwise procedure makes the dependency structure explicit. Spatial localization is evaluated first, semantic fidelity second, and the latter is conditional on the former. That ordering is central to the interpretability claims made for LGS.

## 4. Interpretability and relation to prior metrics

LGS is presented as more interpretable and more reliable than earlier layout-to-image evaluation schemes because it exposes separate failure modes instead of collapsing them into a single pass/fail statistic [2509.16691]. The DetectIoU term is a continuous spatial score, which avoids the all-or-nothing behavior of AP/AR-style recall and precision. Rather than merely determining whether a box crosses a threshold, it reports how closely the predicted region aligns with the conditioned layout.

The semantic gate is intended to prevent incorrect crops from inflating semantic performance. The paper contrasts this with CropVQA, which can produce false acceptance if a VQA model answers an attribute question correctly on a badly localized crop. It also identifies false rejection through semantic VQA on a poor crop, appearance blindness in spatial-only SAMIoU, and discretization failures in BinaryIoU, which reduces alignment to a binary IoU pass/fail criterion [2509.16691].

An additional stated property is open-set evaluation. Because GroundingDINO is used to query each instance description directly, LGS can evaluate arbitrary novel object names or fine-grained captions rather than relying on a fixed taxonomy. The attribute-level decomposition into color, texture, and shape further makes the score diagnostic: it can reveal whether a method places the object correctly but misses appearance attributes, or captures appearance while failing to respect layout. This suggests that LGS is intended not only as a leaderboard metric but also as an analysis tool for model debugging and ablation.

## 5. Integration into DenseLayout

DenseLayout is the benchmark context in which LGS is operationalized. It is described as a dense test set of 5,000 images with approximately 18 instances per image, constructed by generating images with Flux.1-Dev on SA-1B prompts resized to 1024px, tagging open-set categories via RAM++, detecting boxes and labels via GroundingDINO with $\texttt{box\_th}=0.35$ and $\texttt{text\_th}=0.25$, generating fine-grained captions per box with Qwen2.5-VL, and filtering to images with at least 15 instances [2509.16691].

At evaluation time, a model receives the same global prompt together with the list of conditioned instance pairs $(b_i, c_i)$, then generates an image. The LGS pipeline is run on that generated image and the condition list, and the resulting 4-tuple is reported alongside global image-quality metrics: VQA-Score, PickScore, and CLIPScore [2509.16691].

The benchmark role of LGS is therefore twofold. First, it provides a standardized grounding-oriented evaluation on very dense scenes. Second, it complements global quality metrics that may reflect overall visual plausibility or text-image alignment without resolving whether individual conditioned instances are spatially and semantically correct. A plausible implication is that DenseLayout uses LGS to shift evaluation toward instance-level controllability under high object counts.

## 6. Reported empirical behavior

The paper reports DenseLayout results for models built on Flux.1-Dev and SD3-Medium, with the spatial component listed as mIoU and the semantic components listed separately for color, texture, and shape [2509.16691]. On Flux.1-Dev, Regional-Flux obtains mIoU $=14.06$, color $=11.34$, texture $=11.91$, and shape $=11.84$; RAG obtains mIoU $=17.23$, color $=14.22$, texture $=14.62$, and shape $=14.55$; and InstanceAssemble obtains mIoU $=43.42$, color $=27.60$, texture $=29.50$, and shape $=29.14$.

On SD3-Medium, InstanceDiff obtains mIoU $=47.31$, color $=29.48$, texture $=33.36$, and shape $=32.43$; MIGC obtains mIoU $=34.39$; HICO obtains mIoU $=22.42$; CreatiLayout obtains mIoU $=15.54$; and InstanceAssemble obtains mIoU $=52.07$, color $=33.77$, texture $=36.21$, and shape $=35.81$ [2509.16691]. The incomplete attribute values for MIGC, HICO, and CreatiLayout are reported in the source tables but are not reproduced in the supplied data block.

Ablation results on DenseLayout show a monotonic increase in spatial mIoU as components are added: the base system with no Assemble, no cascade, no LoRA, and no DenseSample yields mIoU $=11.69$; adding Assemble-Attn yields $43.98$; adding the cascaded design yields $45.96$; adding LoRA training yields $51.28$; and adding DenseSample yields $52.07$ [2509.16691]. The paper presents these ablations as evidence that LGS tracks model improvements as architectural and training choices change.

The reported failure-mode analysis in Figure 6 further supports that interpretation. CropVQA false acceptance, SAMIoU ignoring semantics, and BinaryIoU discretization failures are presented as cases that LGS resolves through its continuous spatial score and gated semantic checks [2509.16691]. Within the terms of the paper, LGS is thus positioned as a metric that is simultaneously sensitive, diagnostic, and suitable for dense open-set layout grounding.

Source: https://www.emergentmind.com/topics/layout-grounding-score-lgs