A poster or product ad has two parts. One is a background photo with a subject that must not be covered. The other is text that has to sit in the quiet, empty parts of that photo. Current systems build these two parts separately. Layout generators decide where boxes go and never see the words that will fill them. Ad-copy generators write text from a product page and never see the picture. This note argues that the hand-off between them is a hidden source of factual error. When the empty space in a photo is small, the text has to be squeezed to fit. A language model asked to squeeze tends to drop qualifiers, merge facts, and invent short, punchy claims. In other words, a tight space budget turns into a hallucination budget.
The proposal has two parts. First, measure this effect directly: does faithfulness fall as the available space shrinks? Second, test a fix in which the text and the layout are chosen together. The fix retrieves the product facts, splits them into small atomic claims, and computes how many characters each candidate box can hold legibly without covering salient pixels. It then selects whole claims that fit, instead of paraphrasing the facts into a shorter form. A colleague could summarize it this way: "Fitting text into the empty parts of an image pressures the model to make things up. Pick which true facts to show based on how much room there is, rather than compressing all of them."
Core Claim
The central hypothesis is empirical and falsifiable. For an abstractive generator, entity-level faithfulness decreases as the character budget shrinks. Here the budget is set by the low-saliency area of the background, meaning the regions a saliency detector marks as visually unimportant. Replacing compression with budgeted claim selection, coupled to the choice of layout, keeps faithfulness roughly constant across budgets. The cost is that fewer claims are covered, and that cost falls gradually rather than collapsing.
The second hypothesis is that the joint choice matters most on cluttered backgrounds. PosterLayout reports that in these images saliency maps mark most of the canvas as critical. On such images, picking a layout for its text capacity, not only for its visual plausibility, should recover informativeness that a fixed layout loses.
If a simple length-capped prompt reproduces both effects, the method claim fails. The measurement of space-induced hallucination would still stand as a finding.
Where This Sits
Three lines of work motivate the gap.
Content-aware layout models predict boxes as (category, x, y, w, h) tokens conditioned on the image and its saliency map, and they treat text as an afterthought. These include CGL-GAN, PosterLayout's DS-GAN, and RALF. RALF supplies a method we reuse: it retrieves reference layouts by visual similarity and injects them through cross-attention. We use it as the generator that proposes candidate layouts. We do not change its retrieval.
ReLayout names "structural hallucinations" in layouts produced by multimodal language models and fixes them with relation reasoning over salient blocks. Its hallucinations are geometric. It never asks whether the words in the boxes are true, and that is the question we ask.
FaithCAMERA supplies both the textual evidence and a key constraint. It finds that 77.1% of human reference ads contain ungrounded content. It also shows that filtering training data raises entity faithfulness from 68.6 to 79.0 while lowering ROUGE-L. Faithfulness and catchiness therefore already trade off, and we expect a space budget to sharpen that trade.
Two papers constrain the design.
SmartOverlays places labels by minimizing occlusion of salient regions. When labels are too large for the free space, it can only drop or truncate them, because it cannot change the text. That failure is exactly what claim selection is meant to make principled.
Hallucination Begins Where Saliency Drops detects hallucination from internal attention-gradient "saliency." That is a different notion from pixel saliency. It costs 30–40% extra latency, so we use it only as an optional diagnostic, not as part of the method.
Retrieval here is plain text retrieval of product facts. We do not claim novelty in retrieval itself. SKE-Layout warns that generic embeddings ignore geometry. We avoid that problem by computing geometric capacity explicitly instead of retrieving it.
Research Program
Benchmark. We build a paired benchmark, because no existing dataset pairs backgrounds, boxes, and grounded facts. Backgrounds and reference layouts come from PKU PosterLayout and the CGL-Dataset. Source facts come from FaithCAMERA's landing-page inputs and its human-verified faithful references. We pair each canvas with a landing page of a matching product category where categories overlap. The pairing is synthetic, and we will report results separately on category-matched and random pairings so this choice is visible.
Capacity. For each candidate box $b$, the capacity $C(b)$ is the largest number of characters that renders legibly in $b$. Legibly means two things: a font size at or above a minimum $s_{\min}$, and rendered glyph pixels overlapping the saliency map by at most a threshold $\tau$. We obtain $C(b)$ by actually rendering text, not by estimating from box area.
Joint selection. RALF proposes $K$ candidate layouts. A language model splits the retrieved landing-page text into atomic claims $c$, each with a rendered length $\ell(c)$ and an importance weight $w(c)$. The importance weight comes from how often that claim's entities appear in faithful FaithCAMERA references for similar products. We then solve
$$\max_{L \in \{L_1,\dots,L_K\},\; S \subseteq \text{claims}} \sum_{c \in S} w(c) \quad \text{s.t.} \quad \sum_{c \in S_b} \ell(c) \le C(b)\ \text{for every text box } b \in L,$$
where $S_b$ is the set of claims assigned to box $b$. This is a small multiple-knapsack problem, which is NP-hard in general. It is tractable here because a layout has only a few text boxes and a page has only tens of claims, so exact search is feasible.
Light rewriting. Finally, the language model may rewrite the selected claims only lightly, for fluency. An entity-preservation check rejects any rewrite that adds an entity not in the source. Compression, in the sense of shortening by paraphrase, is never requested.
Measuring the effect. The analysis traces faithfulness against budget. We sweep $s_{\min}$ and $\tau$ to vary the available characters on the same canvases. For each budget we measure FaithCAMERA-style entity precision for every system. The prediction is a falling curve for abstractive baselines and a flat curve for selection. We also count specific error types the file's taxonomy predicts under compression. These are dropped hedges, which turn tentative findings into definitive facts (Tang et al.), and invented discounts of the kind FaithCAMERA documents.
Baselines
Naive reference. The do-nothing reference is the decoupled pipeline. A retrieval-grounded language model writes ad copy from the landing page with no knowledge of the image. RALF then places boxes, and the text is rendered into them at whatever font size fits, shrinking below $s_{\min}$ if necessary. This shows how bad the current hand-off is, measured jointly on legibility, occlusion, and faithfulness.
Strongest competitor. The budget-prompted pipeline is the competitor I expect to be hardest to beat. We compute $C(b)$ for RALF's single best layout and tell the same language model "write at most $C(b)$ characters per box, using only facts from this page." This is cheap, uses the same capacity computation, and differs from our method only in compressing rather than selecting. It is also the cleanest ablation of the key ingredient. Both systems get the same language model, the same retrieved passages, the same RALF checkpoint, and the same number of tuning trials on a held-out split; for the baseline, those trials go to prompt variants. Beating it means higher entity precision at matched claim coverage in the tightest third of budgets.
Strongest layout baseline. We include RALF and ReLayout as existing layout methods. Each fills its boxes using the budget-prompted text, so the comparison is on layouts alone. The file reports RALF at FID 3.45 on PKU against a 13.59 baseline (RALF). It reports ReLayout at Overlap 0.0086 against PosterLlama's 0.0211 (ReLayout). Our layouts come from RALF candidates, so we require that layout metrics do not degrade by more than seed variance relative to RALF's top-1 choice while text capacity improves.
Direct multimodal prompting. A strong multimodal language model receives the image, retrieved references, and the landing page, and outputs boxes plus copy as JSON. This is the "just ask a big model" baseline.
Fixed-layout ablation. Claim selection runs on RALF's top-1 layout only. This isolates the contribution of choosing the layout jointly with the text.
Reference numbers. For text faithfulness, the reference number is FaithCAMERA's entity precision of 79.0 for filtered mT5-base. That number has no space constraint. I found no reported result on Lacuna for any system evaluated jointly on saliency occlusion and textual faithfulness, so that comparison has no prior figure.
Why not just…?
Why not just tell the language model the character limit? This is the real threat, and the budget-prompted baseline exists to settle it. The skeptical reading is that modern models obey length limits well and paraphrase faithfully. If the budget-prompted curve of entity precision against budget is as flat as the selection curve, the method contribution collapses. We would then report the benchmark, and the absence of space-induced hallucination when budgets are explicit, as the result. FaithCAMERA's observation that faithfulness and informativeness trade off even without a length constraint is the reason to expect compression to hurt, but this is an expectation, not evidence.
Why not just place the text afterwards with an occlusion-minimizing optimizer? SmartOverlays shows what happens when text cannot change: oversized labels are dropped or truncated, at roughly 4.34 seconds per frame. CGL-GAN reports that decoupled placement such as SmartText underperforms composition-aware generators on subject occlusion and smoothness. Truncation is a factual failure of its own: a cut-off claim can change meaning. The decoupled baseline measures this directly through the rate of truncated or illegible renders.
Why not just prompt GPT-4o with the image and references? ReLayout finds that direct coordinate generation by a multimodal model overlaps badly, 0.0211 versus 0.0086. MRAG-Bench finds that GPT-4o gains only 5.82% from ground-truth retrieved images, while humans gain 33.16%, and that open models often do worse with retrieval than without. Both results suggest the model will neither respect the saliency map nor stay close to the evidence. Still, these results are on other tasks, so we include the system rather than assume it fails.
Why not just generate freely and filter out unfaithful or occluding outputs? On cluttered canvases, unconditioned generators repeatedly propose layouts that hit salient regions, so rejection rates are high. Gradient-based rejection costs 30–40% extra latency (Hallucination Begins Where Saliency Drops). We run a matched-compute rejection baseline: sample until our method's wall-clock budget is spent, then keep the best sample that passes both filters. We report its pass rate across budgets.
Proposed Experiments
Experiment 1: the budget sweep. On roughly 900 PKU test canvases and a CGL test subset, we vary $s_{\min}$ and $\tau$ to span budgets from generous to severe. For every system we report four groups of measures:
- Text faithfulness: entity precision, scored against the landing page with FaithCAMERA's procedure, plus human spot-checks.
- Claim coverage: weighted recall of reference claims.
- Layout metrics: overlap, alignment, and saliency occlusion ($R_{sub}$, $R_{shm}$, $R_{com}$).
- Rendered-text checks: the rendered-text saliency overlap, and the fraction of renders below $s_{\min}$.
The decisive check is the slope of entity precision against budget. The space-induced-hallucination claim requires a negative slope for the budget-prompted and direct-prompting baselines that is significant across seeds. The method claim requires a selection slope near zero, with coverage within a small margin of the budget-prompted baseline.
Experiment 2: clutter stratification. We split canvases by the fraction of pixels above a saliency threshold, then compare joint selection to the fixed-layout ablation within each stratum. Joint choice should help most in the most cluttered stratum. If it does not, the layout-coupling part of the proposal is unnecessary, and we would report that selection on a fixed layout suffices.
Experiment 3: human study. Annotators, who do not know which system produced each poster, judge whether each rendered poster states anything not supported by the landing page, and whether it is legible and attractive. This guards against the entity metric missing hallucinations that are subtle paraphrases.
As an optional diagnostic, we will check whether the internal attention-gradient saliency drop reported by Hallucination Begins Where Saliency Drops occurs more often under tight budgets. That would link the two notions of saliency, but no conclusion of this note depends on it.
References
- PosterLayout: A New Benchmark and Approach for Content-Aware Visual-Textual Presentation Layout
- Composition-aware Graphic Layout GAN for Visual-Textual Presentation Designs
- Retrieval-Augmented Layout Transformer for Content-Aware Layout Generation
- ReLayout: Integrating Relation Reasoning for Content-aware Layout Generation with Multi-modal Large Language Models
- FaithCAMERA: Construction of a Faithful Dataset for Ad Text Generation
- SmartOverlays: A Visual Saliency Driven Label Placement for Intelligent Human-Computer Interfaces
- Hallucination Begins Where Saliency Drops
- SKE-Layout: Spatial Knowledge Enhanced Layout Generation with LLMs
- Evaluating large language models on medical evidence summarization
- MRAG-Bench: Vision-Centric Evaluation for Retrieval-Augmented Multimodal Models