img2img task cards

Paired 1024×1024 input card → output card examples for training image-to-image models to solve a task that is rendered entirely in pixels. The input card holds the question, the choices, the picture and an empty answer area; the output card is the same card with the answer written, marked or pasted in. No text prompt is needed at inference: the task is the image.

This repository is updated in place. There is one dataset, Icey444/img2img; every change to the templates or the data is applied to the files here and recorded in the version table below. The same card lives in the code repository as docs/hf_dataset_card.md.

Versions

version	date	change
v2.1	2026-10-17	Decorative words removed from the skins: no "Exercise" running head and no "Problem." prefix on latex_doc, no "Q." prefix on worksheet. Only task-bearing text is rendered. exam and vqa configs re-rendered.
v2	2026-10-17	First release of the template-v2 render of the 43k bundle. One font size per card; layout engine picks row/column placement and font to maximise readable pixels; answer slot sized from the answer. Multiple-choice answers are "easy first": letter only, or a red circle / green check mark on the correct option's letter, with an explicit instruction. Images stored inline, raw source picture kept as source_image.

Contents

43,020 pairs in 13 categories (one config each), 2,000 of them in a seeded, subsource-stratified validation split.

config	subsources	pairs
abstract_reasoning	arc, jigsaw, pgm, raven_center, raven_dist4, vcog_cvr, vcog_marsvqa	6,387
analogy	vasr	1,300
cognition	corecog	148
document	latex_render, websight	2,700
drawing	fscoco_sketch, quickdraw	2,900
editing	ip2p	1,500
exam	exam_ai2d, exam_mathvista, exam_mmmu, exam_scienceqa, examsv, handwriting_math, latex_math	5,819
physics_video	clevrer, vimeo_next_frame	3,000
puzzle_solving	maze, sudoku	2,200
restoration	colorization, deblur, denoise, inpainting, lowlight, rotation_fix, superres	2,331
simulation	vbvr	235
text_to_image	caption_coco	1,500
vqa	aokvqa, chartqa, docvqa, gqa, ocrvqa, pathvqa, stvqa, tallyqa, textvqa, visual7w, vqa_llava	13,000

from datasets import load_dataset
ds = load_dataset("Icey444/img2img", "vqa", split="train")
row = ds[0]
row["input"], row["output"]          # PIL images, 1024x1024
row["answer_bbox"]                   # where the output differs from the input

Columns

column	meaning
input / output	the 1024×1024 input card and target card (PIL)
source_image	the raw picture placed on the card (max side 1024); None for prompt-only cards
question_text, options, instruction_text	the texts rendered on the input card
answer_text	text answer as written on the output card (for MCQ: the letter)
answer_full	full MCQ option (B. mushroom) when answer_text is a letter
answer_mode	text, mcq_letter, mcq_circle, mcq_check, full_image, in_place
answer_bbox	{left, top, width, height} of the region where output differs from input (slot, mark, or pasted image)
card_template	skin: box, underline, ruled, worksheet, chat, flashcard, latex_doc, prompt_plain, prompt_chat, prompt_min, curated, curated_inplace, or a legacy page render
layout, font_px, text_rect, image_rect, option_cols	layout-engine decision (text_only / column / row / row_slot; rects are [l, t, r, b], empty when not applicable)
source_dataset, orig_id, source_split, extra	provenance; extra is a JSON string with adapter-specific fields
id, subsource, category, task_family	identifiers

Template v2 in one paragraph

One font size per card for question, choices, instruction and answer. For each candidate font the engine tries text-above-image and text-beside-image layouts and keeps the one that maximises (font / 52)^1.5 × image_scale, so neither the text nor the picture is sacrificed. The answer slot is sized from the ground-truth answer and sits at the bottom; it is the answer_bbox, and the output card is pixel-identical to the input outside it. Multiple-choice targets never repeat the option text: the card either asks for the letter in a one-line slot or has no slot and the target draws a red circle or a green check mark on the correct letter. Image-answer tasks return either a padded answer card or the input card with the picture region repainted (in_place). Three real-paper subsources (latex_math, handwriting_math, examsv) keep their scanned-page renders. Nothing decorative is drawn: no headers, titles, page numbers or filler words.

Licenses

Each row's source_dataset names the upstream dataset; rows inherit that dataset's license (COCO CC-BY 4.0, ScienceQA CC-BY-NC-SA, MMMU Apache-2.0, etc.). Check the upstream card before commercial use.