Tecci: Tricky Edits of Collected and Curated Images for Instruction Following

TECCI provides 1,934 images paired with 7,550 edit instructions for evaluating multimodal models. The dataset includes two subsets: TECCI-GGIS with 1,404 images and 7,020 automatically generated instructions, and TECCI-IRCS with 530 images and 530 manually written instructions. Created by Google and last updated in May 2026, it is hosted on Hugging Face.

Use Cases

• Benchmarking model performance on image editing tasks based on the provided edit instructions.
• Training or fine-tuning vision-language models on instruction-following based on the described image-edit pairs.
• Analyzing the difference in model performance between automatically generated and manually written instructions.
• Developing evaluation metrics for multimodal reasoning and task completion.

Strengths

• Includes 7,550 total edit instructions across 1,934 images.
• Provides a manually curated subset (TECCI-IRCS) with 530 human-written instructions for higher quality.
• Offers a larger, automatically generated subset (TECCI-GGIS) with 7,020 instructions for scale.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the full combined dataset is not explicitly stated, which may limit suitability assessment.
• The description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

Google

Collection Method

Images are collected and curated, with edit instructions generated automatically (Gemini 3 Pro) and written manually.

Freshness

Last updated 2026-05-30 16:55:21; freshness should be verified.