Steering Generative Models for Accessibility: EasyRead Image Generation

Nicolas Dickenmann; Yanis Merzouki; Sonia Laguna; Thy Nowak-Tran; Emanuele Palumbo; Julia E. Vogt; Gerda Binder

cs.HCMar 14, 2026

Steering Generative Models for Accessibility: EasyRead Image Generation

Authors: Nicolas Dickenmann, Yanis Merzouki, Sonia Laguna, Thy Nowak-Tran, Emanuele Palumbo, Julia E. Vogt, Gerda Binder

Organizations: Department of Computer Science, ETH Zurich, Switzerland · Unicef Digital Impact Division, Valencia, Spain · ETH AI Center, ETH Zurich, Switzerland

Abstract

EasyRead pictograms are simple, visually clear images that represent specific concepts and support comprehension for people with intellectual disabilities, low literacy, or language barriers. The large-scale production of EasyRead content has traditionally been constrained by the cost and expertise required to manually design pictograms. In contrast, automatic generation of such images could significantly reduce production time and cost, enabling broader accessibility across digital and printed materials. However, modern diffusion-based image generation models tend to produce outputs that exhibit excessive visual detail and lack stylistic stability across random seeds, limiting their suitability for clear and consistent pictogram generation. This challenge highlights the need for methods specifically tailored to accessibility-oriented visual content. In this work, we present a unified pipeline for generating EasyRead pictograms by fine-tuning a Stable Diffusion model using LoRA adapters on a curated corpus that combines augmented samples from multiple pictogram datasets. Since EasyRead pictograms lack a unified formal definition, we introduce an EasyRead score to benchmark pictogram quality and consistency. Our results demonstrate that diffusion models can be effectively steered toward producing coherent EasyRead-style images, indicating that generative models can serve as practical tools for scalable and accessible pictogram production.

Figures & tables

EasyRead design principle	Property description	Metric (score)	Metric description
Low visual clutter	Few colors and minimal visual noise	Palette Complexity ( $s_{\text{palette}}$ )	Color simplicity
Simple geometry	Clear shapes with smooth outlines	Edge Density ( $s_{\text{edges}}$ )	Fine-grained line detail
Clear focus	One visually dominant object	Saliency Concentration ( $s_{\text{saliency}}$ )	Concentration of visual attention
Strong separation	Foreground clearly distinct from background	Foreground–Background Contrast ( $s_{\text{contrast}}$ )	Perceptual foreground–background separation
Stable layout	Main content centered and well sized	Centering Error ( $s_{\text{centering}}$ )	Alignment with the image center
Consistent strokes	Uniform and readable outline thickness	Relative Stroke Thickness ( $s_{\text{stroke}}$ )	Outline consistency and readability

Table 1. EasyRead design principles and corresponding metrics.

Figure 1. Our pipeline to generate EasyRead pictograms. Training (Left) : Input data is preprocessed, captioned using BLIP and augmented before fine-tuning Stable Diffusion v1.5 using Low-Rank Adaptation (LoRA). A unique instance token (sks) is used to learn the EasyRead style. Inference (Right) : The model generates new pictograms by combining the learned style token (sks) with a descriptive prompt and specific color constraints to ensure stylistic consistency. Flowchart of the EasyRead pictogram generation pipeline divided into Training and Inference phases. The Training section illustrates sample data diverging into three parallel steps: Image Preprocessing, BLIP (outputting the caption "Two girls and a boy are standing on a crosswalk"), and ARASAAC-only augmentation (outputting "background color: red; skin color: black; hair color: gray"). These three outputs and an Instance Token are using in training LORA Adapters stacked atop Stable Diffusion version 1.5. The Inference section demonstrates an Instance Token, a text prompt ("A teacher writing on a chalkboard"), and color constraints ("background color: white; skin color: white; hair color: red") feeding directly into the identical model stack to output a matching generated pictogram.

Model	ERS $\uparrow$	CLIP $\uparrow$
SD v1.5 (GT)	$0.40\pm 0.07$	$24.33\pm 2.75$
Ours (LoRA)	$\mathbf{0.47\pm 0.06}$	$\mathbf{31.15\pm 2.88}$

Table 2. Quantitative comparisons between our model and the Stable Diffusion v1.5 baseline on the EasyRead score and CLIP similarity score are summarized as mean

\pm

standard deviation over 275 samples.

GT (SD v1.5)	Ours (LoRA)	Global Symbols	Nano Banana Pro	Prompt (without sks token)
				The skyline of New York City at night; background color: black; skin color: black; hair color: blonde
				A mom reading to a happy child; background color: red; skin color: black; hair color: blonde
				A firefighter on a ladder extinguishing a fire; background color: yellow; skin color: white; hair color: brown

Table 3. Qualitative single-shot comparison of baseline Stable Diffusion v1.5, our model and external closed-source models (Global Symbols [7] and Nano Banana Pro [8]) on identical prompts.

Panel of Figure 2 — Figure 2. Qualitative sample of four generations of our model at different seeds with the prompt: The eifeltower in Paris; background color: white; skin color: white; hair color: black . Four generated images of the Eiffel Tower showing variations from the same text prompt. The images are arranged in a 2 by 2 grid. Despite the prompt explicitly requesting a white background, the outputs vary significantly: \begin{itemize} \itemTop-left: A stylized, colorful drawing of the tower situated above a yellow archway against a light blue sky. \itemTop-right: A sketch-style, distant view of the tower at the end of a path with buildings in the background and a blue sky. \itemBottom-left: A drawing of the tower standing on a gray path bordered by bright green grass, set against a plain white sky. \itemBottom-right: A simple black outline of the tower isolated on a pure white background, being the only image that strictly adheres to the background color constraint. \end{itemize}

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Table 4 — Table 4. Examples of the three datasets in our corpus: ARASAAC pictograms, OpenMoji icons, and LDS symbolic illustrations. Although all are simplified drawings, they exhibit clear and systematic stylistic differences in visual abstraction, line weight, and compositional conventions.

Scene Description	Background	Skin	Hair
The skyline of NYC at night	Red	White	Blonde
A person watering a houseplant	Green	Black	Brown
A child brushing their teeth	Blue	Asian	Dark Brown
A doctor listening to heartbeat	Yellow	Mulatto	Gray
A firefighter climbing a ladder	Black	Aztec	Dark Gray
A teacher writing on a chalkboard	White	White	Red

Appendix

Table 7. Validation Dataset prompts and control attributes

Panel of Figure 3 — Figure 3. Four samples at different seeds generated by the closed-source Nano Banana Pro ( Google DeepMind, 2025 ) model with the prompt: A pictogram of the skyline of New York City at night . A 2 by 2 grid displaying four different generated pictograms of the New York City skyline at night: \begin{itemize} \itemTop-left: A black silhouette of the skyline against a dark blue background, featuring a white crescent moon and stars, with the words ``NEW YORK CITY'' written in black text below the buildings. \itemTop-right: A white outline and solid white shapes forming the skyline against a solid black background, with a white crescent moon and stars. \itemBottom-left: A skyline with dark silhouettes and yellow glowing accents, framed within a dark blue circular background containing a yellow crescent moon and stars. \itemBottom-right: A black silhouette and outline of the skyline against a very pale, off-white background, featuring a black crescent moon and stars. \end{itemize}

Prompt (without sks token)	Generation 1	Generation 2	Generation 3	Generation 4	Generation 5
A street market with food stands; background color: red; skin color: mulatto; hair color: black
A person taking a photograph; background color: white; skin color: asian; hair color: red
A soccer player kicking a ball; background color: green; skin color: asian; hair color: darkBrown
A cup of coffee on a table; background color: green; skin color: white; hair color: darkGray
A police officer directing traffic; background color: white; skin color: mulatto; hair color: darkBrown
A doctor listening to a patient’s heartbeat; background color: yellow; skin color: mulatto; hair color: gray

Appendix

Table 8. Further samples generated by our model. Five samples at different seeds for each prompt

Image Generation

Image generation studies models that create or edit images from text, examples, layouts, and other structured controls.

See all 463 papers

Large Language Models are Universal Reasoners for Visual Generation
May 5, 2026Sucheng Ren, Chen Chen, Zhenbang Wang +5Diffusion Model Guidance Image Generation
1Johns Hopkins University · 2Apple
DrawAI: Agentic Benchmark and Workflow for Making Raster Images Editable
Aug 1, 2026Pu Cao, Qingye Kong, Xuedan Yin +5Image Reconstruction
Beijing University of Posts and Telecommunications · Tsinghua University · Beijing Institute of Technology
LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes
Sep 3, 2026Chuyan Chen, Haoxing Chen, Kun Chen +27Diffusion Transformer T2I Diffusion Models
AGI Research Center, Inclusion AI

May 5, 2026J/K move · Enter open · S save

Large Language Models are Universal Reasoners for Visual Generation

Sucheng Ren, Chen Chen, Zhenbang Wang, Liangchen Song+4

1Johns Hopkins University · 2Apple

Diffusion Model Guidance Image Generation Compositional T2I Generation VLM Reasoning

Open paper AskPDF arXiv

Text-to-image generation has advanced rapidly with diffusion models, progressing from CLIP and T5 conditioning to unified systems where a single LLM backbone handles both visual understanding and generation. Despite the architectural unification, these systems frequently fail to faithfully align complex prompts during synthesis, even though they remain highly accurate at verifying whether an image satisfies those same prompts. We formalize this as the understanding-generation gap and propose UniReasoner, a framework that leverages the LLM as a universal reasoner to convert its understanding strength into direct generation guidance. Given a prompt, the LLM first produces a coarse visual draft composed of discrete vision tokens. It then performs a self-critique by evaluating the draft for prompt consistency, producing a grounded textual evaluation that pinpoints what needs to be corrected. Finally, a diffusion model is conditioned jointly on the prompt, the visual draft, and the evaluation, ensuring that generation is guided by explicit corrective signals. Each signal addresses a limitation of the other: the draft provides a concrete, scene-level anchor that reduces under-specification in text-only conditioning, while the evaluation turns verification into grounded, actionable constraints that correct omissions, hallucinations, and relational errors. Experiments show that UniReasoner improves compositional alignment and semantic faithfulness under the same diffusion backbone while maintaining image quality, demonstrating a practical way to exploit LLM reasoning to close the understanding-generation gap.

Abstract

Image Generation

Explore similar work

Large Language Models are Universal Reasoners for Visual Generation

DrawAI: Agentic Benchmark and Workflow for Making Raster Images Editable

LLaDA-Image: Building Strong Image Generators with Fully Open Training Recipes