Diffusion Reward Modeling for Text Rendering: 14,000 Prompts with Generated Latents

14,000 text prompts designed for evaluating text preservation in images generated by diffusion models. Each prompt includes pre-extracted target text and 5 Stable Diffusion 3 generated latents, totaling 70,000 latents. The dataset was created by author 'leffff' and last updated on August 12, 2025.

Use Cases

• Training reward models for text rendering based on the pre-extracted target text and generated latents.
• Benchmarking text preservation in generated images using the provided prompts and latents.
• Fine-tuning diffusion models for better text-to-image synthesis based on the 70,000 generated latents.

Strengths

• Contains 14,000 specifically designed text prompts for text rendering evaluation.
• Includes 70,000 Stable Diffusion 3 generated latents (5 per prompt).
• Prompts have pre-extracted target text for direct evaluation of text preservation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the primary prompt data is unknown, which may limit suitability assessment.
• The dataset's license is unspecified, which may restrict usage.

Provenance

Source

huggingface

Collection Method

Likely generated using Stable Diffusion 3 models based on designed prompts.

Freshness

Last updated 2025-08-12 12:41:50.