GenEscape: Hierarchical Escape Room Generation

• GenEscape is a hierarchical multi-agent framework that synthesizes escape room images with both visual realism and logical solvability.
• It decomposes the puzzle creation into stages—functional design, scene graph reasoning, layout synthesis, and local image editing—to ensure clear affordances and shortcut avoidance.
• Iterative feedback among specialized agents refines symbolic and spatial representations, yielding geospatially aligned images that support a stepwise solution.

Searching arXiv for the specified paper to ground the article in the current record. GenEscape is a hierarchical multi-agent framework for generating 2D escape room puzzle images that are visually realistic while also being logically solid, meaning solvable, free from unintended solutions, and visually clear in terms of affordances. It addresses the difficulty that base image models struggle with spatial relationships and affordance reasoning by decomposing generation into structured stages: functional design, symbolic scene graph reasoning, layout synthesis, and local image editing. Specialized vision-LLM agents collaborate through staged, iterative, and feedback-rich interaction so that the resulting scene is visually coherent and functionally solvable (Shan et al., 27 Jun 2025).

1. Definition and problem setting

GenEscape formulates escape room image generation as a constrained synthesis problem in which visual quality is necessary but insufficient. The target output is an image that supports an intended stepwise solution, blocks shortcuts, and communicates manipulable elements through clear affordances. In this setting, “logically solid” refers specifically to solvability, shortcut avoidance, and affordance clarity, which are treated as first-class requirements rather than byproducts of generic text-to-image prompting (Shan et al., 27 Jun 2025).

The central premise is that one-shot image generation is poorly matched to this objective. The system therefore introduces a hierarchical multi-agent methodology that bridges textual design, symbolic reasoning, low-level layout, and pixel-level image editing. This staged decomposition suggests that escape room synthesis is treated not merely as semantic image generation, but as a structured design-and-verification pipeline in which intermediate representations are explicitly inspected and revised (Shan et al., 27 Jun 2025).

2. Hierarchical decomposition and agent roles

GenEscape decomposes the generation process into four main stages, each handled by specialized agents. The stages are functional design, symbolic scene graph reasoning, layout synthesis, and affordance-guided local image editing (Shan et al., 27 Jun 2025).

At the functional design stage, the Designer Agent receives a scene type, objects, and optionally a solution length, and generates three outputs: a scene description, an initial scene graph, and an official solution. The scene description is a natural language description setting environment and puzzle premise. The initial scene graph is a YAML-formatted hierarchy where nodes are objects and relationships encode spatial or containment dependencies. The official solution is a stepwise textual solution representing valid player actions. This stage establishes an explicit blueprint in which necessary objects and their relationships are actionable rather than implicit (Shan et al., 27 Jun 2025).

At the symbolic scene graph reasoning stage, the Player Agent simulates a human player and proposes a plausible solution by reasoning solely over the scene graph, while the Examiner Agent checks whether the Player’s solution matches the official one. If not, the Examiner points out discrepancies, especially shortcuts, missing links, or unclear logic. The Examiner then refines the scene graph to block shortcuts and reinforce the intended path. This loop repeats until agreement is reached, thereby ensuring that the symbolic structure supports only the intended solution, with no loopholes (Shan et al., 27 Jun 2025).

At the layout synthesis stage, the Builder Agent transforms the refined scene graph into a 2D spatial layout in which each object is rendered as a positioned icon reflecting the specified spatial relationships and constraints. The Player then simulates solution based on the layout, covering visual cues and possible misinterpretations, and the Examiner checks for alignment between perception from the layout and the intended solution. The Builder revises the layout based on design critiques, improving placement, labeling, and clarity. The stated goal is to guarantee visual grounding so that every logical object or connection in the symbolic scene appears visually and is correctly spatialized (Shan et al., 27 Jun 2025).

At the affordance-guided local image editing stage, the Builder converts the layout into a photorealistic image using text-to-image synthesis and then iteratively edits local regions to clarify or suppress object affordances. The Player solves the puzzle by interpreting the rendered image rather than only the layout, the Examiner checks whether the Player’s actions reconstruct the official solution and identifies ambiguous affordances, visual artifacts, or shortcut opportunities, and the Builder refines the image locally. The outcome is intended to make the solution visually obvious to a human, block non-intended solutions, and clarify what can and cannot be manipulated (Shan et al., 27 Jun 2025).

3. Optimization by feedback

The system is organized around a repeated optimization-by-feedback loop applied at the graph, layout, and image stages. For each stage, the Player derives a solution $S^*$ from the current representation $R$, the Examiner compares $S^*$ with the official solution $S$, and the relevant agent refines $R$ until the discrepancy set $\Delta$ is empty (Shan et al., 27 Jun 2025).

The paper summarizes this process as follows:

Input: Initial scene graph G0, official solution S R = G0 for stage t in [Graph, Layout, Image]: repeat: S* = player.solve(R) Δ = examiner.check(S, S*) if t == Graph: R = examiner.refine(R, Δ) elif t == Layout: R = builder.refine(R, Δ) elif t == Image: R = builder.refine(R, Δ, R_L) until Δ is empty Output: Final image R

The corresponding formulation is given as:

$$\begin{align*} \text{Repeat for each stage } t \in \{\text{Graph, Layout, Image}\}:\ \quad \text{Repeat until } \Delta = \emptyset:\ \qquad S^* \gets \mathsf{player}.\mathsf{solve}(R)\ \qquad \Delta \gets \mathsf{examiner}.\mathsf{check}(S, S^*)\ \qquad \begin{cases} R \gets \mathsf{examiner}.\mathsf{refine}(R, \Delta), & \text{if } t = \text{Graph} \ R \gets \mathsf{builder}.\mathsf{refine}(R, \Delta), & \text{if } t = \text{Layout or Image} \end{cases} \end{align*}$$

A key property of this procedure is stage-wise progression control: each feedback loop only advances when the Player’s derived solution matches the official solution according to the Examiner. This establishes a verification condition over multiple modalities rather than relying on the final image alone (Shan et al., 27 Jun 2025).

4. Solvability, shortcut avoidance, and affordance clarity

GenEscape defines its core functional criteria explicitly. Solvability means that at every stage the Player must be able to derive the official solution from the available information, whether graph, layout, or image, and the Examiner explicitly checks for logical errors or impossibilities. Shortcut avoidance means that if the Player finds an unintended solution path, the Examiner notes the exploit and refinement actions such as moving objects, adding blocks, or changing object states are made to restore puzzle integrity. Affordance clarity means that at the image stage local edits such as emphasizing a lock or removing ambiguous routes are made until the Player correctly perceives and selects the intended manipulation (Shan et al., 27 Jun 2025).

These criteria are not treated as separate post hoc evaluation dimensions only; they are embedded into the generation workflow itself. The Player–Examiner loop functions as an internal adversarial check on both symbolic and visual ambiguity. This suggests that GenEscape operationalizes puzzle validity as agreement between intended and inferred action sequences across representations, rather than as a purely semantic similarity objective (Shan et al., 27 Jun 2025).

A common misunderstanding would be to interpret the system as only a visual prompt-engineering pipeline. The reported methodology instead centers on symbolic correction, spatial grounding, and local affordance editing, with image synthesis appearing only after earlier representational inconsistencies have been reduced. Another plausible misconception is that higher semantic similarity to the initial prompt is the primary target. The reported results indicate otherwise, since CLIP score does not necessarily improve because solvability tweaks may visually diverge from the initial design prompt (Shan et al., 27 Jun 2025).

5. Evaluation protocol and empirical results

The paper evaluates GenEscape with human evaluation and reports five metrics: Solvability, Shortcut Avoidance, Spatial Alignment, CLIP Score, and #Gen, the average number of generation iterations required. Solvability is defined as the percentage of images where humans can solve the puzzle as intended. Shortcut Avoidance is the percentage where unintended or trivial solutions are blocked. Spatial Alignment is the percentage where the image visually reflects the relationships in the scene graph. CLIP Score measures semantic similarity between the designer prompt and the generated image. #Gen measures the average number of generation iterations required (Shan et al., 27 Jun 2025).

Method	Solv.	Short.
GPT-4o	3.3%	0.0%
+Desc (D)	6.7%	3.3%
+D+Scene Graph	6.7%	13.3%
+D+S.G.+Layout	10.0%	20.0%
+D+S.G.+Image	20.0%	16.7%
Ours full	53.3%	46.6%

Method	Align.	CLIP	#Gen.
GPT-4o	N/A	N/A	N/A
+Desc (D)	0.0%	0.42	N/A
+D+Scene Graph	26.7%	0.37	N/A
+D+S.G.+Layout	13.3%	0.38	13.2
+D+S.G.+Image	23.3%	0.32	15.8
Ours full	36.7%	0.32	4.5

The reported findings are that hierarchical agent collaboration improves output quality in terms of solvability, shortcut avoidance, and affordance clarity while maintaining visual quality. Scene graph and layout stages especially improve spatial reasoning and solution path clarity, while image refinement and affordance editing are described as crucial for preventing shortcuts and ambiguous object usage. The full system also requires fewer generation iterations than the incremental layout and image variants shown in the table, with #Gen reported as 4.5 for the full method compared with 13.2 and 15.8 for two partial baselines (Shan et al., 27 Jun 2025).

The CLIP results complicate a purely prompt-faithfulness interpretation of success. The full method has a CLIP score of 0.32, lower than some simpler baselines, which the paper explains by noting that solvability tweaks may visually diverge from the initial design prompt. This indicates a tension between functional correctness and prompt-level semantic similarity rather than a simple monotonic trade-off in which all metrics improve simultaneously (Shan et al., 27 Jun 2025).

6. Conceptual significance and design implications

The paper identifies several key insights. First, the modular, stage-wise design moves from highly structured symbolic representations to spatially grounded visuals and finally to pixel-level affordance-aware editing. This progression is presented as better matched to the complex requirements of puzzle design than one-shot image generation. Second, feedback-driven refinement reliably blocks unintended exploits and increases human solvability. Third, the hierarchical agent-based optimization is described as computationally efficient because it requires fewer image generations and enables fine-grained control at each reasoning level (Shan et al., 27 Jun 2025).

In methodological terms, GenEscape exemplifies division of labor among four agents specializing in design, solving, critiquing, and image generation. The collaboration loop is multi-modal: the Player solves in the current modality, the Examiner critiques divergence from the ideal solution, the Designer, Builder, or Examiner make the appropriate edits, and completion occurs only when all agents agree that the scene permits only the intended solution. This suggests a general paradigm in which controllable generation is achieved by alternating synthesis and adversarial verification over progressively richer representations (Shan et al., 27 Jun 2025).

A plausible implication is that GenEscape belongs to a class of systems in which correctness constraints cannot be robustly imposed at the pixel level alone. The use of scene graphs and layouts indicates that symbolic and geometric structure are treated as indispensable intermediate abstractions for aligning perception, interaction logic, and final rendering. Within the scope of the reported paper, this implication remains interpretive, but it is directly supported by the architecture and ablation-style comparisons (Shan et al., 27 Jun 2025).

7. Scope, limitations, and interpretation

GenEscape is specifically framed around generating 2D escape room puzzle images. Its contribution is therefore not a general theory of puzzle solving, nor a generic text-to-image enhancement method, but a task-specific framework for scenes that must be both visually appealing and functionally solvable. The paper’s evaluation emphasizes human solvability, shortcut avoidance, spatial alignment, and affordance clarity, which means performance is assessed on functional criteria critical for puzzle design rather than on aesthetics alone (Shan et al., 27 Jun 2025).

The results do not support the claim that semantic image similarity metrics such as CLIP are sufficient proxies for puzzle quality. On the contrary, the reported findings explicitly note that CLIP score does not necessarily improve under the refinements that increase solvability and shortcut avoidance. This makes GenEscape relevant to a broader discussion in generative modeling: success criteria may depend on whether the generated artifact is intended for passive viewing or for structured interaction (Shan et al., 27 Jun 2025).

The paper also avoids claiming absolute guarantees in a formal verification sense beyond the stated loop structure and agreement condition among agents. Its “guarantees” are described operationally: the symbolic structure is refined until the intended solution is supported and loopholes are removed, visual grounding is checked at the layout stage, and affordance ambiguities are corrected locally at the image stage. This indicates a practical, iterative notion of correctness grounded in agent agreement and empirical evaluation rather than exhaustive proof over all possible human interpretations (Shan et al., 27 Jun 2025).