What makes the best AI thumbnails: v4 experiment (2026-11-04)

Scope:

12 latest videos, 288 ideas (6 ways of finding threads x 4 per video), 120 approved (10 per video).
619 renders across 19 generation recipes, plus 46 helper images (clay plans, face-repair tests, price probes). All through Magnific.
Every recipe render scored blind by two separate judges, plus code checks (face match, brightness, stray text).
Spend: 58,295 credits by my ledger, under the 70,000 cap (see "Credits" below).

The answer: the best combination

Recipe "WIN" (now the default in RUNBOOK_v4.md):

Model: GPT Image 2.5 through Magnific (high quality, 2K).
References, in this order:

- 3 photos of you: a natural expression photo for the emotion, your favorite front photo, a natural three-quarter smile. - 1-2 real photos of the object. - The real-photo character sheet. - One clean round-5 render as a style reference.

Prompt: the long round-5 prompt (the image roles, then SHOT / SUBJECT / OBJECT / SCENE, then the bright-room block).
Layout: let the prompt place everything. Base images, clay plans and 3D plans all scored lower.

Head to head on 85 new ideas, against the round-5 recipe you called your best so far:

WIN beat it on 55 ideas and lost on 28, with 2 ties (sign test p = 0.004). Average gain +0.60 standard units, about half a judge point.
Counting which of the three Stage-2 recipes made each idea's best render (ties split), out of 85 ideas:

Recipe	Best render
WIN	43.5
ALT (round 5 with a curated face photo)	21
Round 5	20.5

What made the biggest difference (screening: 24 ideas, one change at a time from the round-5 recipe)

Each number is the change versus the round-5 recipe, in standard units within an idea. One unit is about 0.8 of a judge point on the 1-10 scale. 90% intervals are in brackets. Bold means the interval excludes zero. Each cell is one render per idea, so effects under about 0.25 are noise.

What I changed	Effect	Face match passes
Model: Nano Banana 2.1	-1.47 (-1.67, -1.26)	4% (round 5: 73%)
Model: Nano Banana Pro	-1.58 (-1.78, -1.35)	4%
Model: Seedream 5.0 Pro	-1.73 (-2.07, -1.37)	33%
Layout: an outlier thumbnail as the base image, edited Marcus-style	-0.55 (-0.87, -0.24)	64%
Layout: AI gray clay plan first, then render from it	-0.98 (-1.24, -0.71)	53%
Layout: Blender 3D gray plan first (9 ideas only)	-0.60 (-0.90, -0.28)	78%
Prompt: short (70 words)	-0.46 (-0.79, -0.13)	77%
Prompt: director style (camera, grid, first/second read, avoid list)	+0.16 (-0.09, +0.41)	75%
Refs: face photo only, no objects or body	+0.03 (-0.22, +0.29)	91%
Refs: 3 face photos instead of 1	+0.06 (-0.16, +0.26)	91%
Refs: character sheet instead of the body panel	+0.21 (+0.04, +0.39), weak	83%
Refs: your fixed favorite photo instead of the emotion-matched one	+0.36 (+0.16, +0.55), not robust (below)	70%
Refs: a style reference (clean round-5 render)	+0.14 (-0.09, +0.34)	87%
Refs: everything (3 photos + sheet + objects + style reference)	+0.33 (+0.13, +0.55)	87%

Robustness check. I re-judged three of the same images (round 5, fixed favorite, everything) in smaller 4-image groups, together with one new render per idea using a curated natural photo:

"Everything" held: +0.45 (+0.07, +0.81).
The fixed favorite fell to -0.03.

So the small reference effects depend on how they are judged. The full bundle holds up, and Stage 2 confirmed it (+0.60).

Ranked by size: 1. The model, by far. 2. Layout aids, which hurt. 3. The full reference bundle. 4. Prompt length.

Your three hunches

1. Reference images (which, how many, how chosen). The full bundle wins (+0.33 screening, +0.45 re-judged, +0.60 head to head). One photo vs three photos alone made no clear difference. - How the face photo is chosen hides a trap. 80 of your 182 expression photos are stares: all 7 "Surprise (Happy)" photos and 18 of the 41 "Happy" ones. - The old rule picked photos by face-match score, which favors wide-open eyes, and GPT copies the stare. That's likely why judges flagged a "bulging wide-eyed stare" in most renders. - A curated natural photo on its own was not clearly better (+0.09 in the 4-image groups, a new render per idea; +0.24 in Stage 2, sign test p = 0.19). Treat it as a fix, not the lever. 2. Prompts. Long and specific beats short (-0.46 for short). The director format was no better than the round-5 prompt. 3. Generation methods. - Character sheet: small plus (+0.21, weak). - 3D plan (gray clay): -0.60. Thin evidence: 9 ideas, one render each, no colour plan, and not your new body model. I did not measure whether it fixed pose, only the overall judged quality; it did keep the face match higher (78%). - AI clay plan and outlier-comp base: both lower (-0.98, -0.55).

How threads were found (6 methods)

APPEAL of each method's ideas, read from the round-5 recipe render, so the recipe is the same for every idea.

Method	Approved	APPEAL (90% CI)
M3 Outlier object twist (outlier objects from the topic and other niches, twisted)	35 of 48	6.09 (5.88, 6.31)
M5 Desire first (stakes, curiosity gap, the moment before)	18 of 48	6.09 (5.75, 6.44)
M1 Title threads (Marcus: pull directions from the title)	25 of 48	5.98 (5.82, 6.14)
M4 Video moment (the real payoff from the video)	12 of 48	5.96 (5.62, 6.33)
M2 Outlier composition first	14 of 48	5.71 (5.33, 6.08)
M6 Channel winners remix (your own 27-154x thumbnails)	16 of 48	5.36 (5.00, 5.75)

What this does and doesn't show:

Only one result is clear: remixing your own past winners was the weakest source. Judges read those ideas as repeats with lower stakes.
Outlier object twist and desire first tie at the top, and their intervals overlap title threads and video moment.
Outlier object twist did produce the most approved ideas (35 of 48).
Only approved ideas were rendered, so this can't show strong ideas the selectors threw away.

How ideas were decided (the decision-making test)

Correlation between each selector's pre-render score and how good the round-5-recipe thumbnail turned out (Spearman ρ within video, 12 videos):

Selector	ρ (90% CI)
Teen viewer simulation ("you are 15, which would you click")	+0.26 (+0.09, +0.41)
Rules judge (Marcus formula + Dill's 4 pillars + your rules)	+0.21 (+0.05, +0.36)
Data score (views and outlier ratio behind the idea)	+0.02

Both judgment selectors predict modestly, with overlapping intervals.
The data score does not predict at all. A strong outlier source does not mean a strong thumbnail.
Consensus picks scored higher than wildcards (6.06 vs 5.64), but that mixes in method: most video-moment ideas came in as wildcards, and most outlier-object ideas by consensus.

Face match and the face fix

Average match: GPT keeps your face at 0.76. About 72% of its renders pass the 0.72 floor; Nano Banana and Seedream pass 4-33%.
AI re-edits of the face did not raise the match:

| Face repair | Average match | |---|---| | None (raw render) | 0.678 | | GPT edit | 0.646 | | Nano Banana edit | 0.670 |

The real-photo paste ran on 5 of 9 test renders (its pose and lighting gates blocked 4). It raised the match to 0.76-0.83 on 4 of them and dropped one to 0.575.
On the finals it wiped out the expression on all 6 I checked: the grimace, the mud-slap wince and the noodle slurp went blank, and one broke the scene (facefix_real_paste_check.jpg). I kept the AI faces. That matches your round-5 call: expression fit beats a likeness number.
28 of the 120 finals sit under the 0.72 floor. They are flagged on the page.

What GPT refuses

GPT Image refused about 1 in 10 ideas: fire on people, real-looking guns, blades near faces, a school bottle that read as drinking. Those 7 ideas were rendered with the same recipe on Nano Banana 2.1, and their faces are weaker (0.63 average match).

Defects the judges keep naming (next fixes)

Bulky arms on you (likely the body panel and character sheet).
Small faces in wide outdoor scenes.
Mud and water spray that look like round beads.
Molten metal and lava that read as water or gel.
Made-up lettering on props.
Charred bill edges that kill "it doesn't burn" twists.
Objects that look store-bought instead of 3D printed.

What this test did not cover

Your verdict. The judges agree with your past picks 65% of the time (81% without 3 renders carrying flaws you named; that cut was made after seeing the results). Your blind picks on the pick page are the real test: paste "Copy my picks" back and the recipes get rescored on your choices.
Ideas the selectors rejected were never rendered.
No new photos of you. Only the existing 182-photo expression library was used.
Model fairness. The 3 non-GPT models got the GPT-tuned prompt, with no prompt written for them.
Comp-as-base, clay and 3D were each tested with one approach per idea.

Credits

My ledger (logs/credits.jsonl) says 58,295. Refused GPT renders are logged at 0, which was not verified.
The whole wallet fell 66,425 over the same hours: 394,145 read at 00:17 EST, before the first v4 call at 00:27, and 327,720 at the end. That drop also includes another thumbnail session rendering at the same time, so my true spend is at most 66,425. Under the cap either way.
Seedream's price was never measured cleanly; it is logged at 150 per image.