Checkpoint 1 · R4b hint-image samples

Hint images, locally generated: pick the winners

Your R4 ruling asked: can we generate good, style-consistent hint images fully locally, for free? Below are 28 real samples across 5 vocabulary cards — pick the winners, and decide whether to go further.

6 image sets · 28 candidates Z-Image Turbo · offline · Apache-2.0 ~45s/image on this Mac Samples only — nothing in production
Checkpoints Latest →
What these are for

The hint pop-up, illustrated

Plain-English framing before the picture galleries.

Every flashcard's hint pop-up carries two mnemonic "scenes" — short emoji stories meant to help you remember the word (we call these slots 4 and 5). Right now those scenes are text-only. The idea tested here: turn each scene into a small illustration that shows up right next to the text, using a model that runs entirely on this laptop, for free, with a license that's fine to ship commercially.

Five real vocabulary cards were used to test this — panas, kerja, jalan, mata, bawa — with the actual mnemonic scene text pulled straight from each card's real hint data. For each scene, 4 candidate images were generated from the same prompt with 4 different random seeds, so you can judge both the concept and the variety.

Set 1 of 6

panas — "tense, critical"

Scene (slots 4 & 5 share one illustration): a pineapple and a race car in a tense negotiation — 🍍 + 🏎️.

All four keep the flat-vector storybook look and land the "tense negotiation" idea reasonably well. Seed 101 has an odd robot-like car that reads a little off; 202 and 404 are ENR's picks as the cleanest reads.

panas-s45 candidate seed 101
seed 101
ENR suggestion panas-s45 candidate seed 202
seed 202
panas-s45 candidate seed 303
seed 303
ENR suggestion panas-s45 candidate seed 404
seed 404
Set 2 of 6

kerja — "celebration, party"

Scene (slots 4 & 5 share one illustration): a wedding — 🎭 (masked dance) + 🍮 (dessert) — 4 candidates.

303 centers the bride most explicitly — ENR's top pick — with 202 as runner-up. All four are usable; this is a "which reads clearest" call more than a "which is broken" call.

kerja-s45 candidate seed 101
seed 101
ENR suggestion kerja-s45 candidate seed 202
seed 202
ENR suggestion kerja-s45 candidate seed 303
seed 303
kerja-s45 candidate seed 404
seed 404
Set 3 of 6

jalan — "intermediary, connector"

Scene (slots 4 & 5 share one illustration): a strawberry jam jar next to a ladder joint — 🍓 + 🪜 — shown as two batches, v1 then v2.

This is the one honest miss, and the fix. v1 (seeds 101–404) drew the ladder standing straight up instead of bridging two things, and the jam read dark/ambiguous — it didn't land the "connector" idea. v2 (seeds 505–808) used one prompt fix — spelled the ladder out as a horizontal bridge, added "bright glossy" and a literal strawberry prop — and it worked. Same model, same 45 seconds/image, one iteration.

v1 — first pass (the miss)
jalan-s45 candidate seed 101
seed 101
jalan-s45 candidate seed 202
seed 202
jalan-s45 candidate seed 303
seed 303
jalan-s45 candidate seed 404
seed 404
v2 — after the prompt fix
ENR suggestion jalan-s45 candidate seed 505
seed 505
jalan-s45 candidate seed 606
seed 606
jalan-s45 candidate seed 707
seed 707
ENR suggestion jalan-s45 candidate seed 808
seed 808
Set 4 of 6

mata — "eye of the storm"

Scene (slots 4 & 5 share one illustration): a map and a tooth caught in a storm's calm center — 🗺️ + 🦷.

Minimalist across all four; the storm swirl itself is subtle against the warm color palette — worth a look at whether that reads clearly enough at hint-popup size. ENR's picks: 101 and 303.

ENR suggestion mata-s45 candidate seed 101
seed 101
mata-s45 candidate seed 202
seed 202
ENR suggestion mata-s45 candidate seed 303
seed 303
mata-s45 candidate seed 404
seed 404
Set 5 of 6

bawa — "cause, result in" (slot 4)

Scene (slot 4 only — this card's aid judge flagged this scene FAIL): a bat bounces off a wall, unharmed.

Worth flagging: the model draws the bricks visibly breaking in all four, despite "unharmed" being in the prompt — a real semantic tension between what the words say and what the model draws. ENR's picks (101, 404) are the least dramatic about it, but none of the four fully avoid it.

ENR suggestion bawa-s4 candidate seed 101
seed 101
bawa-s4 candidate seed 202
seed 202
bawa-s4 candidate seed 303
seed 303
ENR suggestion bawa-s4 candidate seed 404
seed 404
Set 6 of 6

bawa — "cause, result in" (slot 5)

Scene (slot 5 — a different story from slot 4, same card): a bat sets off a wall, which sets off a barrel — a chain reaction.

Amusing side effect: the model turned "chain reaction" into a literal chain in all four images. Harmless, arguably charming, and consistent across every seed. ENR's picks: 101, 404.

ENR suggestion bawa-s5 candidate seed 101
seed 101
bawa-s5 candidate seed 202
seed 202
bawa-s5 candidate seed 303
seed 303
ENR suggestion bawa-s5 candidate seed 404
seed 404
Style

Where the consistency comes from — and its ceiling

Every image above shares its look — flat vector, warm palette, soft rounded shapes, single clean scene — from one fixed style phrase appended to every prompt, nothing more. That's already carrying the whole corpus's visual consistency by itself. A trained style LoRA (a small custom style model, fit to a handful of curated winners) would lock that consistency in harder and reduce set-to-set drift — that's the "phase 2" ask below.

Decisions ledger

What was decided vs. what's implied

1
Generate hint images fully locally, for free, with a commercial-safe license — feasible?
Yes — this run is the proof. Decided by Frank in the R4 ruling that set this test up.
Decided by Frank
2
Use hand-picked real cards (not synthetic test words) so the test reflects production content.
Five real cards were used — panas, kerja, jalan, mata, bawa — with real mnemonic scene text.
Inferred from what you said
3
Don't touch production data until Frank explicitly approves — samples stay samples.
Nothing was written to emoji_image_url or any other production field; all 28 images live only in this review page and on local disk.
Inferred from what you said
Worth your eye

Three calls to make

1
Approve the overall look and direction? Flat-vector storybook style, warm palette, one clean scene per illustration — is this the right visual language for hint images?
2
Train a style LoRA on the winners for corpus-wide consistency (phase 2)? Locks the look in harder than the style phrase alone — a bigger one-time investment (dataset curation + a training run) for consistency at scale.
3
Wire the winning picks into emoji_image_url for these 5 cards as a live pilot? Currently samples only. A "go" here writes real image URLs onto real cards for the first time — small, reversible, but a first production step.