Hand-Quality Labelling · v0.31

(Last edited · 04 Aug 2025)

Why this matters Every accurate label you create nudges our generative-image models closer to anatomically believable hands. Thank you for being part of that.


1 Purpose

Modern diffusion and transformer models still struggle with hands: wrong finger counts, melted geometry, odd proportions. We’re building a high-quality dataset—cropped hand images with human judgements of “passable vs. defective.”

Your labels will:

  1. Train automatic defect detectors (pre-filtering bad hands).

  2. Feed research into anatomy-aware generation.

Short version – judge the hand, not the whole artwork; judge severity, not resolution.


2 Rating Design

2.1 Questions (unchanged)

Q#PromptAllowed answersNotes
1How would you rate the quality of this hand?Good, Bad, Not Sure, No HandsNot Sure ≤ 5 % overall.
2If Bad, why?Severely Deformed · Missing⁄Extra Digits · Subtle Distortion · OtherMulti-select disabled—pick the single worst defect you see.

2.2 Ground Rules

  • Severity first. A tiny, far-away hand that is clearly fused is still Severely Deformed.

  • One bad hand → bad image. If any visible hand fails, the whole crop is Bad.

  • Ask, don’t guess. When torn, choose Not Sure and ping @Yada in hands-labelling.


3 Rating Examples – Quick Lookup

CategoryConcise definitionImage placeholders
GoodFive digits, plausible shape, no obvious artefacts.GOOD-01 …
Bad — Severely DeformedFused, melted, broken, impossible joints.SEVERE-01 …
Bad — Missing⁄Extra DigitsLooks human, but digit count ≠ 5 or ghost digits.DIGIT-01 …
Bad — Subtle DistortionFive digits yet odd proportions / “noodle” fingers.SUBTLE-01 …
Bad — OtherRare defects that fit none of the above.OTHER-01
Not SureAmbiguous after a second look.NS-01
No HandsNo hand visible at all.NOHAND-01 …

4 Rating Workflow

[Flow-chart to be inserted] — text version below.

  1. Hands present? • No → No Hands • Yes → Step 2

  2. Severe deformity visibly present? • Yes → Bad – Severely Deformed

  3. Digit count off? • Yes → Bad – Missing⁄Extra Digits

  4. Five digits yet anatomy “off”? • Yes → Bad – Subtle Distortion

  5. OtherwiseGood

  6. Still unsure at any pointNot Sure and leave a comment.

(Tip: zoom; don’t upscale. If blur alone hides the anatomy, choose Not Sure.)


5 Ambiguity Handling – Mini FAQ

ScenarioGuidance
Partial hand occludedIf visible part shows a clear defect → label Bad; if hidden → Not Sure.
Stylised / cartoon handsAsk whether stylisation keeps five distinct, proportionate digits. If yes, possibly Good.
Multiple hands in cropJudge the worst hand.
Motion blur / low-resBlur hides anatomy → Not Sure. Defect still obvious → relevant Bad category.
Edge of frameCropped finger counts as “missing” only if clearly absent, not merely out-of-frame.

(Add visual edge-case call-outs when available.)


6 Examples of Past Mis-Labelling

“Crisp ≠ correct.” In the pilot, many crisp, painterly hands were marked Good despite elongated or fused fingers. Remember: looks-human ≠ anatomically-sound.

Common errorCorrect labelWhy
Sharp, five-digit hand but fingers twice normal lengthBad – Subtle DistortionProportions matter.
Human-drawn but fused thumb-indexBad – Severely DeformedFusion overrides artistry.
Cartoon mitt with three fingersNot Sure (or Bad – Other if policy set)Lacks standard anatomy; flag.

(Insert illustrative crops.)


7 Version History

VerDateChanges
0.113 Jul 2025Initial clean-up.
0.227 Jul 2025Added quick table; trimmed prose.
0.301 Aug 2025Pilot feedback; ambiguity notes.
0.4-scratch04 Aug 2025Polished structure; workflow, mini FAQ, mis-label table; placeholders added.

Next actions

  1. Fill blanks: insert flow-chart and final image examples.

  2. Run a 100-image calibration batch—track where Not Sure is used.

  3. Update FAQ / examples, then bump to v0.4 final.

Questions, edge cases, or uncertainty? → Post in hands-labelling or tag @Yada.