KE-Check: Knowledge-Intensive Image Framework

• KE-Check is a two-stage framework that enhances scientific fidelity in text-to-image generation by ensuring domain-specific constraints are met.
• Its knowledge elaboration stage converts terse prompts into structured, curriculum-aligned descriptions for accurate diagram construction.
• The checklist-guided refinement phase audits and edits images with atomic, verifiable rules to correct visual inaccuracies and underlying errors.

KE-Check is a two-stage framework for knowledge-intensive text-to-image generation that aims to improve scientific fidelity rather than only perceptual realism or coarse instruction following. It was introduced together with KVBench in "Knowledge Visualization: A Benchmark and Method for Knowledge-Intensive Text-to-Image Generation" (Zhao et al., 24 Apr 2026). In this setting, the central problem is that knowledge visualization requires not only semantic alignment but also strict adherence to domain knowledge, structural constraints, and symbolic conventions, creating a gap between visual plausibility and scientific correctness. KE-Check addresses that gap through Knowledge Elaboration for structured prompt enrichment and Checklist-Guided Refinement for explicit constraint enforcement through violation identification and constraint-guided editing.

1. Problem setting and benchmark context

Knowledge-intensive visualization is distinguished from ordinary natural image generation by the requirement that generated images comply with scientific and curricular constraints. The motivating failures are not merely aesthetic defects: existing text-to-image models can omit required components, assign incorrect labels, or render physically incorrect arrow directions. The examples given include missing components in a cell diagram and incorrect force-arrow orientations in a physics sketch (Zhao et al., 24 Apr 2026).

The framework is introduced in the context of KVBench, a curriculum-grounded benchmark for evaluating knowledge-intensive text-to-image generation. KVBench covers six senior high-school subjects: Biology, Chemistry, Geography, History, Mathematics, and Physics. The benchmark consists of 1,800 expert-curated prompts derived from over 30 authoritative textbooks; the detailed description further specifies 1,800 bilingual prompts in English and Chinese, with each sample paired with a reference image and a 4–6-item checklist (Zhao et al., 24 Apr 2026).

A concise summary of the benchmark configuration is given below.

Aspect	Specification
Subjects	Biology, Chemistry, Geography, History, Mathematics, Physics
Prompt set	1,800 expert-curated bilingual prompts
Source material	Over 30 authoritative textbooks
Per-sample annotation	Reference image and a 4–6-item checklist
Evaluation tracks	Brief Caption; Detailed Caption
Model pool	14 state-of-the-art open- and closed-source models

The benchmark is used to show substantial deficiencies in logical reasoning, symbolic precision, and multilingual robustness, with open-source models consistently underperforming proprietary systems. Within that empirical setting, KE-Check is proposed as a method for mitigating scientific hallucinations and narrowing the performance gap between open-source and leading closed-source models (Zhao et al., 24 Apr 2026).

2. Two-stage architecture

KE-Check is explicitly described as a two-stage framework. Its design premise is that no prior method jointly injects rich domain context into prompts and systematically enforces a fine-grained, verifiable checklist of constraints. KE-Check fills this gap by decoupling knowledge enrichment from explicit constraint auditing and editing (Zhao et al., 24 Apr 2026).

The two stages are as follows.

Stage	Function	Output
Knowledge Elaboration	Convert a terse prompt into a structured, domain-rich description	Elaborated description $d$ and initial image $I_0$
Checklist-Guided Refinement	Construct, audit, and enforce atomic constraints	Refined image $I^*$

This decomposition is technically significant because it separates latent knowledge supplementation from post-generation constraint enforcement. A plausible implication is that the method is intended to reduce errors arising both from underspecified prompts and from failures of visual realization after a strong prompt has already been supplied.

3. Knowledge Elaboration

The first stage takes a terse prompt $x$ and converts it into a structured, domain-rich description $d$ that encodes three types of information: core scientific concepts and component hierarchy; spatial layout rules, such as panel arrangement and foreground/background; and visual specifications, such as line types, color schemes, and diagram style (Zhao et al., 24 Apr 2026).

The elaboration step is implemented by prompting an MLLM with a scientific system prompt that guides it to expand $x$ under high-school curriculum conventions. Formally, the output is written as

$d = \operatorname{Elab}(x).$

A diffusion- or transformer-based text-to-image generator $G$ then produces the initial image

$I_{0} \;=\; G\bigl(d\bigr).$

The role of this stage is not merely to lengthen the prompt. It structures the prompt around domain content, spatial organization, and visual standards. This suggests that KE-Check treats prompt enrichment as a controlled translation from a terse linguistic request into a curriculum-aligned specification for diagrammatic generation.

4. Checklist-Guided Refinement

After generating $I_0$, KE-Check enforces compliance with an explicit set of atomic constraints. These constraints are derived from the elaborated description $I_0$0 and are formulated as declarative, binary-verifiable rules. If the structured checklist is denoted

$I_0$1

then each item $I_0$2 expresses a specific condition such as the presence of an arrow, the orientation of a symbol, or the inclusion of a required label (Zhao et al., 24 Apr 2026).

The item-wise auditing stage assigns a violation indicator

$I_0$3

and collects the violation set

$I_0$4

The violation set is then aggregated into non-redundant editing instructions $I_0$5. These instructions, together with the initial image $I_0$6, are passed to an image-editing module, either in-context with the same text-to-image model or with a dedicated editor, to correct only the non-compliant regions. The final output is

$I_0$7

which is described as satisfying all $I_0$8 (Zhao et al., 24 Apr 2026).

The pipeline is summarized by the following pseudocode given for KE-Check:

$d$1

The method therefore operates as an explicit generate-audit-edit loop, but with the checklist grounded in the elaborated description rather than introduced independently.

5. Formalization and evaluation protocol

KE-Check is accompanied by a checklist-based evaluation formalism. The constraint set for a prompt $I_0$9 is written as

$I^*$0

For an image $I^*$1, the violation set is

$I^*$2

Each checklist item receives a binary score

$I^*$3

The sample-level checklist accuracy is

$I^*$4

Over $I^*$5 samples, the overall benchmark score is

$I^*$6

This evaluation is conducted on two tracks: Brief Caption, described as implicit knowledge, and Detailed Caption, described as explicit constraints (Zhao et al., 24 Apr 2026). The checklist-based metric is central because it directly measures whether atomic scientific requirements are satisfied, rather than whether the output appears globally plausible.

A recurrent misconception in text-to-image evaluation is that semantic alignment alone is a sufficient indicator of correctness. The formulation used here rejects that premise: compliance is assessed item by item against structural, symbolic, and labeling constraints. In this sense, the benchmark operationalizes scientific correctness as checklist satisfiability rather than holistic similarity.

6. Empirical results and representative examples

On the Brief Caption track for Qwen-Image, the reported checklist accuracy without KE-Check is $I^*$7 and $I^*$8. With KE-Check, the scores become $I^*$9 and $x$0, corresponding to absolute gains of $x$1 in Chinese and $x$2 in English (Zhao et al., 24 Apr 2026).

The ablation study further separates the effects of the two components:

Configuration	Chinese	English
Baseline	20.09	25.74
+Knowledge Elaboration	36.17	28.12
+Refinement w/o Checklist	46.50	30.69
KE-Check (full)	47.67	30.24

These values indicate that both knowledge enrichment and refinement matter, while the full method combines them through explicit checklist construction. This suggests that KE-Check is not reducible to prompt expansion alone.

Two representative examples illustrate the mechanism. In the free-body diagram of a block on a $x$3 incline, the initial image $x$4 misses the normal-force arrow, mis-orients the weight arrow, and omits the angle label. The checklist contains: “There is an arrow labeled ‘mg’ pointing vertically downward”; “There is an arrow labeled ‘N’ perpendicular to the plane”; and “The incline angle ‘30°’ is marked between plane and horizontal.” The detected violations are $x$5 and $x$6, and the resulting editing instructions are: “Add a normal-force arrow labeled ‘N’ at the contact point, draw the 30° angle annotation between the incline and horizontal base, keep existing ‘mg’ arrow.” The final output $x$7 correctly shows both arrows and the angle label (Zhao et al., 24 Apr 2026).

In the chloroplast example, the initial image omits the grana stacks, draws thylakoid membranes incorrectly, and lacks the label “stroma.” The checklist requires visibility of the double outer membrane, depiction of grana as stacked discs, light shading of the stroma region, and labels for the major sub-components. The reported violations are $x$8, $x$9, and $d$0. The editing instruction is: “Insert stacks of thylakoid discs labeled ‘grana’, add light shading and label ‘stroma’, preserve correct membranes.” The result is described as a textbook-accurate chloroplast diagram (Zhao et al., 24 Apr 2026).

Taken together, these examples show that KE-Check targets failures at the level of atomic scientific constraints. The method is therefore best understood not as a generic image-quality enhancement procedure, but as a framework for enforcing curriculum-grounded, binary-verifiable correctness in knowledge visualization.