ShahirHF/eg2-vision-gen2

eg2-vision gen2: synthetic watercolour children's-book pictures

80,000 synthetic pictures in a children's-book watercolour style, each with the exact prompt and seed that made it. They were made as teacher data for the eg2-vision project (24 Oct 2026). 20,000 prompts × 4 seeds.

All pictures are machine-generated. There are no real photographs in this dataset. Prompts often mention people; they are drawn as watercolour illustrations.

Contents

path	what
images/images-NN.tar	PNG pictures, 512 px, 5,000 per tar (WebDataset-style tars). File {i:06d}.png is row i of the prompts file. Rows 0–79,999.
prompts.jsonl	one row per picture, 80,000 rows
prompt_lists.json	the word lists used to build the non-caption prompts
gen_mode.json	generation mode n512s1: native 512 px, 1 step
logs/, score/, look/, code/	generation logs, evaluation scores, sample pages and the scripts that made the data

Prompt fields: i, prompt_id, k (seed index), seed, src (coco, localized_narratives or built), scene, prompt, held_out, char, setting. Held out: the 1,000 rows with held_out: true (250 prompts × 4 seeds).

How the pictures were made

• Model: SANA-Sprint 1.6B (Efficient-Large-Model/Sana_Sprint_1.6B_1024px_diffusers), bf16, 512 px with resolution binning off, 1 step, one CUDA generator per image seeded with seed.
• Prompt = "a children's book watercolor illustration of " + scene.
• The starting noise can be rebuilt with code/sana.py: noise_for("n512s1", seed).

Licences and sources

This dataset combines material under several licences. By using it you agree to all of them.

• Pictures: generated with SANA-Sprint 1.6B, released under the Apache License 2.0. SANA-Sprint's text encoder is Gemma-2-2B-IT, provided under the Gemma Terms of Use; uses must follow the Gemma Prohibited Use Policy.
• Prompts:
  • 12,000 of 20,000 prompts are COCO 2017 train captions (CC BY 4.0), lower-cased, one sentence, 5–14 words.
  • The other 8,000 are built from word lists written for this project (prompt_lists.json).
• Only caption text was used from these sources, never their images.

Citations

• COCO: Lin et al., "Microsoft COCO: Common Objects in Context", ECCV 2014. Annotations CC BY 4.0, https://cocodataset.org.
• Localized Narratives: Pont-Tuset et al., "Connecting Vision and Language with Localized Narratives", ECCV 2020. CC BY 4.0, https://google.github.io/localized-narratives/.
• SANA-Sprint: Chen et al., "SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation", 2025.

Related sets

ShahirHF/eg2-vision-gen2, -gen3 and -gen4 together hold 450,000 pictures (rows 0–449,999 of gen4's prompts.jsonl). ShahirHF/eg2-vision-teacher (FLUX.2 [klein] 4B) and ShahirHF/eg2-vision-ming-wc768 (Ming-Image-0.1-Design) hold pictures from other teacher models.