Comi AI · Hard-Negatives
Image review portal · v1

Your task: tell us which photos are real, which are wrong.

We're teaching the Comi AI scanner to tell apart Colombian foods that look similar (like Casabe vs arepa). For 19 confused pairs, we already pulled 2,455 candidate photos from Google. You don't search anything. You just look at each photo and tap Approve, Reject, or Skip.

19
Confused pairs
2,455
Photos pre-pulled
2–4 hrs
Full queue (~30–45 min / 500 photos)

How it works — 5 simple steps

1

Type your name once

On the next screen, type your first name (e.g. angel). Every approve/reject you make gets tagged with that name so we can track your work and pay you. You only do this once per session — your name stays saved.

2

Pick a pair from the top bar

The top of the screen has 19 buttons, one per pair. Each pair shows two foods like Casabe vs arepa. Pick one to start.

What you'll see
Casabe vs arepa

HN-008-casabe-vs-arepa · failures: 1

Target: CasabeNeed: 30 photos
Confuser: arepaNeed: 50 photos
3

Each pair has 2 sides — review both

Below the pair name there are two side tabs: Target (the food the scanner SHOULD say) and Confuser (the food it currently confuses with). Each side has its own photos. Review the side that has photos waiting, then click the other side tab.

4

For each photo: tap Approve, Reject, or Skip

  • Approve — the photo clearly shows the food on this side. Real food, not edited, not a logo, not a menu, not a collage.
  • Reject — the photo is wrong. A reason chip appears below — pick the reason. (11 reasons listed below.)
  • Skip — you genuinely can't tell. Don't guess. Skip it.

Always click the source link first (under each photo) to check it's a real food page, not a thumbnail collage or expired domain.

5

When you're done, tap Download CSV at the bottom

The portal saves every decision automatically. When you've worked through all 19 pairs (or whenever you stop for the night), tap Download CSV. It downloads a file like comi-hard-negatives-yourname-2026-04-28.csv. Email that file back to Victor.

How to decide: Approve vs Reject

✓ APPROVE only if ALL these are true

If you're unsure, REJECT. Better to reject 5 maybes than approve 1 wrong photo.

✗ REJECT if ANY of these are true

Pick one of these 11 reasons:

wrong_foodPhoto shows a different food than this side's label.
multiple_dishesPhoto shows several different dishes or a mixed plate. Use this for the exact case Angel asked about.
duplicateYou already approved this exact photo.
menu_or_textMenu screenshot, ad, or photo with text/recipe-graphic overlay.
collageMultiple foods stitched together. We need single-dish photos.
low_qualityBlurry, dark, sideways, too small, partially blocked.
ai_generatedLooks AI-made (DALL-E, Midjourney, weird symmetry, "too perfect").
stock_photoWatermarked stock-photo preview from Shutterstock / Getty / etc.
watermarkLogo or watermark covers the food.
not_photoDrawing, illustration, vector, clipart — not a real photo.
inaccessibleSource page won't load, image broken, or expired domain.

Important rules

Ignore the small gray website tag in the top-right corner of each photo (e.g. periodicodelmeta.com). That's part of the review tool showing you the source website — it is NOT part of the photo itself. Don't reject a photo for "having text on it" just because of that tag.
Use the EXACT label shown on the side tab. If the tab says arepa, the label is arepa — not "arepas", not "Arepas", not "arepa con queso". The label is what gets exported in the CSV. Don't change it.
Always open the source page by tapping the link under the photo. Image-search thumbnails are not enough proof — the source page is what tells you if it's a real food photo or a collage / stock site / dead link.
Your work saves automatically. If you close the tab, your name and decisions stay saved. Re-open the portal, click Continue reviewing, and pick up where you left off.

What happens after you submit

Victor runs an automatic validator on your CSV. It checks every row: is the pair_id real, does the label match the pair, is the URL valid, is the reject reason allowed, are there duplicates. If validation passes, your approved photos go into Comi AI's training data so the scanner learns the difference. You'll be paid based on the count of approved + valid rejected rows.

Start reviewing →
Goal: ~672 approved photos across all 19 pairs (50 per food class). Expect 2 to 4 hours for the full queue, or 30 to 45 minutes per 500 photos — easily splittable across sessions, your progress saves automatically. Pay rate confirmed with Victor before you start.
Questions? Text Victor at +1-818-266-7784 (English or Spanish). For technical issues with the portal, screenshot what you see and send.