AI thumbnail readability test: check the text at 168 by 94 pixels
Before you publish an AI thumbnail, shrink it to a small display size and read the text. If it fails there, regenerate with fewer words and larger type.
Before you publish an AI-made thumbnail, shrink it to a small display size, about 168 by 94 pixels, and read the words with your eyes half closed. If you cannot, regenerate with fewer words, larger type and stronger contrast. Most thumbnails are viewed small, and a 16:9 image that is fine at full size can turn into noise.
Why test small
The model draws the image at its output size, and you judge it at that size. Viewers see a tiny version beside others. The test is a downscale, so it costs nothing and catches the common faults: thin letters, low contrast, three-line headlines and a face too small to read.
A script for the check
Download the result, shrink it and put it next to a few of your older thumbnails so you judge it in context.
from PIL import Image
img = Image.open("thumb.png").convert("RGB")
small = img.resize((168, 94), Image.LANCZOS)
sheet = Image.new("RGB", (168 * 3 + 40, 94 + 20), "white")
for i in range(3):
sheet.paste(small, (10 + i * 178, 10))
sheet.save("thumb-small-check.png")What to change when it fails
- Cut the headline to two to four words.
- Ask for a solid color block behind the text for contrast.
- Ask the text to fill at least a third of the frame width.
- Move the subject to one side so the text has a clean area.
- Generate on a text-strong model and keep
qualityhigh.
Generate a thumbnail built for the test
Ask for few words, a plain background behind the text and a large subject. The ratio is 16:9, which every Sume image model listed for text work supports.
curl -X POST https://api.sume.com/v1/images \
-H "Authorization: Bearer $SUME_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"ideogram/ideogram-v4.5","prompt":"YouTube-style thumbnail, the words FIX IT FAST in huge white block letters on a solid red band, a surprised person on the right, bright background","aspect_ratio":"16:9","quality":"medium"}'Test variants, then publish
Use n: 4 for four takes, shrink all of them into one sheet and choose by legibility, not by beauty at full size. Spell-check the text by eye before upload, and keep the winning prompt and job id. For related checks, see Ideogram 4.5 size source.
Related posts
More in Use cases
- All-hands recap with an AI presenter: a 60-second video and its cost
A weekly company recap read by a Sume avatar: script structure, the 60-second cap, a cost table for one week and a year of 52, and the disclosure to add.
- Amazon will flag AI-generated people to shoppers: what to plan
Amazon says it will add a shopper indicator where an image includes AI-generated people, in all stores. Here is what is known, what is open, and how to plan.
- Amazon's AI-people tag: which images are exempt
Amazon wants contains-synthetic-performer on photorealistic AI people. Real people (even AI-altered), characters, no-people and non-photoreal images are exempt.
- Amazon BFCM inventory arrival dates: which SKUs get sale clips first
Amazon's BFCM inbound dates are Oct 14, 21 and 28 by route. Make clips first for SKUs that arrive by Oct 14. Read 2026-10-05, with a queue sketch.
Written by Sume