id stringlengths 16 16 | task_type stringlengths 17 26 | scene_description stringlengths 129 342 | expected_physics stringlengths 55 145 | observed_error stringlengths 74 172 | deliverable_text stringlengths 128 312 | deliverable_files stringlengths 76 137 | tags stringlengths 37 67 |
|---|---|---|---|---|---|---|---|
cube_letters_001 | text_in_image_constraint | 6 cubes in a 2x3 layout on a neutral tabletop. Each cube shows exactly 3 visible faces. Cubes are numbered 1–6 (small number near the base). Each cube must have exactly three different letters, one per visible face, following strict alphabet groups: (1) A/B/C, (2) D/E/F, (3) G/H/I, (4) J/K/L, (5) M/N/O, (6) P/Q/R. No o... | All letters must be legible, non-duplicated within each cube, and match the per-cube mapping. Exactly 18 visible faces total (6 cubes × 3 faces). | Model duplicates letters and omits others; common failures: H missing in cube 3, L missing in cube 4, O missing in cube 5; letters swap across cubes or repeat on two faces. | PASS if the image contains exactly 6 cubes, each with 3 visible faces, numbered 1–6, and the visible letters match: cube1={A,B,C}, cube2={D,E,F}, cube3={G,H,I}, cube4={J,K,L}, cube5={M,N,O}, cube6={P,Q,R}. FAIL on any missing letter, duplicate letter on a cube, wrong grouping, extra letters, or unreadable text. | ["reference_files/cubes_expected.png", "reference_files/cubes_observed.png"] | counting,hard_constraint,text_rendering,cube_faces,alphabet_mapping |
cube_letters_002 | minimal_pair_regression | Recreate the same cube scene as cube_letters_001, but with slightly different camera angle (still 3 faces per cube) and consistent serif font. Keep the same strict letter mapping and numbering 1–6. | The mapping must remain correct under angle variation; text must remain consistent and legible. | Angle change increases hallucinations: cube 2 shows D/E/E, cube 6 shows P/R/R or Q replaced; numbering sometimes disappears. | PASS if mapping and numbering are preserved under the new angle: cube1 A/B/C, cube2 D/E/F, cube3 G/H/I, cube4 J/K/L, cube5 M/N/O, cube6 P/Q/R, with all 6 numbers visible. FAIL otherwise. | ["reference_files/cubes_expected_angle2.png", "reference_files/cubes_observed_angle2.png"] | regression,angle_variation,text_rendering,consistency |
cube_letters_003 | layout_strictness | Same requirement as cube_letters_001, but enforce exact grid placement: top row cubes 1–3 left-to-right, bottom row cubes 4–6 left-to-right. Numbers must correspond to position. | Spatial layout must match numbering; mapping must match letters per cube. | Correct letters but cubes appear shuffled (e.g., cube 6 placed where cube 2 should be) or numbers mismatch position. | PASS if both constraints hold: (a) positions correspond to numbers 1–6 in reading order, (b) letters per cube match the required groups. FAIL if either position/numbering or letter mapping is wrong. | ["reference_files/cubes_expected_grid.png", "reference_files/cubes_observed_grid.png"] | layout,ordering,numbering,hard_constraint |
cube_letters_004 | text_legibility_gate | Same as cube_letters_001 but with 'engraved ink' style: letters carved/printed into the cube surface. Keep high contrast and sharp edges so every letter is readable. | All 18 letters are readable at 100% zoom; no ambiguous shapes (e.g., H looking like N). | Letters become stylized/ambiguous, making verification impossible; some faces blur or merge strokes. | PASS if every letter is unambiguous and the cube mapping is correct. FAIL if any letter is illegible/ambiguous OR mapping is wrong. | ["reference_files/cubes_expected_legible.png", "reference_files/cubes_observed_legible.png"] | legibility,ocr_like_check,text_in_image,verification |
cube_letters_005 | negative_constraint | Same cube scene as cube_letters_001. Additionally: forbid any other text, watermark, captions, or extra symbols anywhere in the image (including background). Only the letters and numbers are allowed. | No extraneous text beyond required letters and numbers. | Model adds decorative words, texture text, or stray letters in background. | PASS if only allowed text appears: A–R on cube faces and digits 1–6 near cube bases. FAIL if any extra characters appear anywhere. | ["reference_files/cubes_expected_noextra.png", "reference_files/cubes_observed_noextra.png"] | negative_prompt,spurious_text,strict_filter |
cube_letters_006 | multi_step_reasoning | Storyboard requirement: panel 1 shows empty tabletop. Panel 2 shows placing cube 1 with A/B/C. Panel 3 adds cube 2 with D/E/F. Continue until panel 6 shows all cubes 1–6. Each panel must preserve earlier cubes unchanged. | Across panels, previously placed cubes retain correct letters; no drift between panels. | Between panels letters drift, duplicates appear, or earlier cubes change when new cubes are added. | PASS if every panel preserves already-correct cubes and each new cube introduced matches required letter group and numbering. FAIL if any panel changes earlier cubes or breaks mapping. | ["reference_files/storyboard_expected_6panels.png", "reference_files/storyboard_observed_6panels.png"] | storyboard,temporal_consistency,multi_panel,planning |
cube_letters_007 | counterfactual_check | Generate two images: (A) the correct cube mapping as in cube_letters_001, (B) an intentionally incorrect mapping where cube 3 swaps H with I (cube 3 shows G/I/I). Label images 'A' and 'B' in the corner. | System must follow instruction that only image B is wrong and image A is correct; text labels A/B must be present and correct. | Both images end up wrong, or labels A/B missing or swapped; model fails to keep one correct while making the other wrong on purpose. | PASS if image A is fully correct per mapping AND image B contains exactly the specified error (cube 3 has G/I/I) with all other cubes correct; A/B labels must be present. FAIL otherwise. | ["reference_files/pair_expected_A_correct_B_wrong.png", "reference_files/pair_observed.png"] | minimal_pair,control_condition,contrastive_eval |
cube_letters_008 | robustness_to_style | Same mapping as cube_letters_001 but rendered in 3 different styles: (1) photorealistic, (2) pencil sketch, (3) ink + parchment. Keep letters and numbers identical across styles. | Style changes must not change the underlying discrete constraints (letters, numbering, layout). | In stylized versions the mapping breaks and letters drift (e.g., cube 6 becomes P/R/R or Q becomes O). | PASS if all 3 styles preserve identical discrete constraints (letters+numbers+layout). FAIL if any style breaks mapping or text. | ["reference_files/triple_style_expected.png", "reference_files/triple_style_observed.png"] | style_robustness,domain_shift,text_rendering |
cube_letters_009 | counting_plus_text | Extend mapping to 8 cubes in a 2x4 layout, numbered 1–8. Letter groups: 1=A/B/C, 2=D/E/F, 3=G/H/I, 4=J/K/L, 5=M/N/O, 6=P/Q/R, 7=S/T/U, 8=V/W/X. Exactly 8 cubes, 3 faces visible each. | Exact cube count (8) and correct per-cube letter groupings; no missing or duplicated letters within any cube. | Model outputs wrong number of cubes, merges cubes, or fails text constraints (missing T, duplicating W, etc.). | PASS if there are exactly 8 cubes numbered 1–8 and each cube shows exactly its 3-letter group (A–X in order by groups). FAIL on wrong count, wrong numbering, missing/duplicate letters, or extra text. | ["reference_files/cubes8_expected.png", "reference_files/cubes8_observed.png"] | scaling,counting,text_constraints,hard_numeric |
cube_letters_010 | verification_prompt_effect | Same as cube_letters_001 but add an explicit instruction inside the prompt: 'Before finalizing, verify cube-by-cube that all 18 faces match the mapping. Do not output if any mismatch.' | Stronger self-verification instruction should reduce mapping mistakes. | Despite verification instruction, output still contains missing letters or duplicates; indicates weak internal checking. | PASS if mapping is correct (same criteria as cube_letters_001). Record whether verification instruction changed outcome compared to baseline. FAIL if any mismatch remains. | ["reference_files/baseline_observed.png", "reference_files/with_verification_observed.png"] | self_check,planning_vs_rendering,eval_prompting |
cube_letters_011 | image_edit_preservation | Image-edit task: given a correct baseline image, edit ONLY cube 5 by changing its material (e.g., marble→wood) while preserving all letters and numbering. No changes to any letters or other cubes. | Edits must preserve discrete text constraints; only texture/material changes allowed. | Model alters letters during edit (e.g., N becomes M) or changes other cubes unintentionally. | PASS if only cube 5 material changes and all letters/numbers remain identical across all cubes. FAIL if any letter/number changes or other cubes are modified. | ["reference_files/edit_input_correct.png", "reference_files/edit_expected_only_material_change.png", "reference_files/edit_observed.png"] | image_edit,attribute_preservation,text_stability |
cube_letters_012 | ocr_alignment_check | Same as cube_letters_001, but letters must be axis-aligned and centered on each face, with consistent font size across all cubes. | Text placement consistency enables automatic checking (OCR/heuristics). | Letters vary in scale/rotation; some are off-center or skewed; increases error rate and makes evaluation noisy. | PASS if mapping is correct AND letters are consistently centered, upright, and similarly sized on all faces. FAIL if mapping breaks or text placement is inconsistent enough to hinder verification. | ["reference_files/aligned_expected.png", "reference_files/aligned_observed.png"] | alignment,consistency,ocr,measurement |
Cube Text Constraint Evaluation Dataset (GDPval)
Overview
This dataset provides a focused collection of human-designed evaluation cases for assessing strict text-to-image constraint satisfaction in multimodal generative models.
The tasks are intentionally simple in visual appearance but logically rigid, requiring precise planning, global consistency, and explicit verification.
The dataset is designed to expose repeatable failure modes where models appear to understand instructions but fail during execution.
This work is intended for model evaluation and qualitative benchmarking, not for training.
Motivation
Multimodal image generation models frequently fail not due to unclear instructions, but due to:
- insufficient global reasoning,
- lack of systematic verification,
- planning shortcuts during single-pass generation.
This dataset isolates those weaknesses by enforcing exact symbolic constraints that are easy for humans to verify but challenging for models to execute correctly.
Task Description
Each task requires generating exactly six cubes, each with three visible faces, labeled with unique letters following a strict alphabetical mapping:
| Cube | Required Letters |
|---|---|
| 1 | A, B, C |
| 2 | D, E, F |
| 3 | G, H, I |
| 4 | J, K, L |
| 5 | M, N, O |
| 6 | P, Q, R |
Hard Constraints
- Exactly three different letters per cube
- No duplicated letters across cubes
- No missing letters
- One letter per visible face
- Global verification across all 18 faces
Any deviation (duplication, omission, misplacement) constitutes a failure.
What This Dataset Evaluates
The dataset targets the following evaluation dimensions:
- Constraint satisfaction under strict symbolic rules
- Global planning vs local rendering behavior
- Multi-object consistency
- Systematic verification vs assumption-based generation
- Repeatable failure modes under clarified instructions
Notably, the dataset demonstrates cases where models repeat the same logical error even after clarification, highlighting weaknesses in internal verification rather than instruction understanding.
Dataset Structure
Each row represents a single evaluation case and includes:
id– unique task identifiertask_description– full task specificationexpected_behavior– precise correctness criteriaobserved_failure– common failure patterndeliverable_text– concise expected outcome description (GDPval requirement)deliverable_files– optional references (e.g. comparison images)
Reference images (when present) are illustrative only and are not used for training.
Intended Use
This dataset is designed to support:
- GDPval-style human-in-the-loop evaluation
- Comparative benchmarking across multimodal models
- Analysis of structural reasoning reliability
- Identification of planning vs execution gaps
It is particularly suitable for evaluating models such as:
- text-to-image generators,
- multimodal reasoning systems,
- image generation pipelines with symbolic constraints.
Not a Training Dataset
⚠️ This dataset is NOT intended for model training.
It is strictly designed for:
- evaluation,
- analysis,
- qualitative benchmarking,
- failure mode documentation.
License
This dataset is released under the Creative Commons Attribution 4.0 (CC BY 4.0) license.
You are free to use, share, and adapt the dataset with proper attribution.
Acknowledgements
Created as part of an independent effort to improve reliability, transparency, and trustworthiness in multimodal generative systems through structured evaluation rather than anecdotal examples.
- Downloads last month
- 11