Tap any image to see it full size (tap again to zoom to actual pixels).

3D process vs v4 (the Studio recipe)

On your sketch, adding the 3D plan made v4 follow your composition: match 6 out of 10 without it, 8 with it. The overall thumbnail scores were close (7 to 7.5). The best single image came from the 3D process with a color plan, which kept the drink pressed up with air at the mouth in all 4 images.

Best image (blind judge's pick)

Best image (blind judge): 3D process, color plan

3D process, color plan. The judge: the biggest, most saturated glass, the drink pressed against the coaster with a visible air gap, the chain in your fist, the right tilt. Its flaws: a half-smile rather than full shock, and some empty gray at the bottom.

Same concept, four ways

All four used the same plan text (your sketch written out) and GPT Image 2.5 at high quality.

v4 WIN: exactly what the Studio runs: 3 face photos, 2 coaster photos, character sheet, style anchor, STYLE block, 2K.
WIN + 3D: the same, plus my real-pose 3D plan (gray) as the layout reference.
3D color / 3D gray: last round's runs: 1 face photo, 2 coaster photos, body panel, the 3D plan, 1K.

Version	Correct	Thumbnail	Match to your sketch	Natural body	Real photo	Likeness
v4 WIN (what the Studio runs now)	4/4	7.0	6.0	7.0	6.5	0.8
v4 WIN + the new 3D plan	4/4	7.0	8.0	8.0	8.0	0.8
3D process, color plan	4/4	7.5	8.0	8.0	7.5	0.78
3D process, gray plan	4/4	7.0	7.5	8.0	8.0	0.8

Medians of 4, all 16 judged together, blind, by an AI against your sketch (green tag = glass upside down, hanging from the chain in your fist). Every image passed that check and your likeness bar.

v4 WIN (what the Studio runs now): 4/4 correct

v4 WIN (what the Studio runs now), image 1: thumbnail 7/10

v4 WIN (what the Studio runs now), image 2: thumbnail 7/10

v4 WIN (what the Studio runs now), image 3: thumbnail 7/10

v4 WIN (what the Studio runs now), image 4: thumbnail 6/10

v4 WIN + the new 3D plan: 4/4 correct

v4 WIN + the new 3D plan, image 1: thumbnail 7/10

v4 WIN + the new 3D plan, image 2: thumbnail 7/10

v4 WIN + the new 3D plan, image 3: thumbnail 7/10

v4 WIN + the new 3D plan, image 4: thumbnail 7/10

3D process, color plan: 4/4 correct

3D process, color plan, image 1: thumbnail 7/10

3D process, color plan, image 2: thumbnail 8/10

3D process, color plan, image 3: thumbnail 7/10

3D process, color plan, image 4: thumbnail 8/10

3D process, gray plan: 4/4 correct

3D process, gray plan, image 1: thumbnail 7/10

3D process, gray plan, image 2: thumbnail 7/10

3D process, gray plan, image 3: thumbnail 7/10

3D process, gray plan, image 4: thumbnail 7/10

What the 3D adds

Your composition: without the 3D, v4 drifted. In all 4 the glass tilted the wrong way, and one came out smaller with empty space. With the 3D plan the tilt and the chain mostly land where you drew them (one 3D-guided image put the fist and chain lower than planned). Both v4 versions still touched or cut the fist at the top edge in 3 of 4.
The chain: two of the four v4-only images got the coaster's chain wrong (toy beads, a bucket handle). None of the 3D-guided images had those chain errors, though two 3D color images got the cord loop wrong (a ring on his finger, no loop in the fist). The coaster itself still went wrong now and then either way (a flipped cork top in one v4 and one v4 + 3D image, a lid-like or oversized coaster in two 3D-process images).
A natural body: 8 vs 7.
The drink: the gray plans made it worse. A gray plan can't show where the drink sits, and with one the drink filled the glass to the mouth (reads like a lidded jar) in 3 of 4 (v4 + 3D) and 4 of 4 (3D gray), vs 1 of 4 for v4 alone. The color plan shows the drink pressed up with air at the mouth, and all 4 color images kept that air gap.
The expression: a grin or half-smile instead of shock in 4 of 4 v4 and 3 of 4 v4 + 3D, vs 1 of 4 in each 3D-process version. I don't know why yet: the 3D-process runs used a smiling face photo too, and the 3D doesn't set the face.

The two processes

	v4 WIN	3D process
Steps	Card, then 7 references, then GPT Image	Card, then a pose photo, then SAM 3D Body and Blender build the frame, then GPT Image with the frame as a reference
Composition	The prompt places things, so it varies	Camera, pose and object positions fixed in 3D
Cost for 4 images	about 360 credits	about 410 (the 360 plus one 50-credit pose photo)
Time	about a minute	about 10 to 30 minutes of my blocking per concept, then a minute
Where	Studio	M5 (SAM 3D Body + retarget), then any Mac

How much to trust this

One concept, 4 images per version, and an AI judge (on your past picks it agrees with you about 65% of the time).
The v4 experiment's earlier result (3D plan -0.60) used the old gray MakeHuman plan on 9 other ideas; it doesn't contradict this, it's a different 3D.
My 3D runs were 1K, v4 2K. The judge saw them all at about the same size (1200 px wide).

Cost of this comparison: 8 v4 renders at 90 credits (720; the v4 ledger is now 139,875 of your 150,000 cap). The 3D rows are last round's images, not re-made. Files: /Users/hafumedia/Desktop/Claude/work/in-development/thumbnail-system/v2/blocking3d/runs/q04s_v4cmp (all 16 + judge), Studio run /Users/hafumedia/Desktop/Claude/work/in-development/thumbnail-system/v4/runs/drinksketch_compare.

Your call

Add the 3D plan, in color, to the v4 recipe as an optional step for concepts where the composition, pose or physics matters (everything else stays v4)?