GPT Image quality test on your approved thumbnails

Same prompt and same reference photos as the 5 thumbnails you approved for A/B tests, made again with GPT Image 2.5 at high (90 credits, what we use now), medium (60) and low (15). Every image is the same 2K size: quality is how much work the model puts into each picture, not how many pixels it has. Two fresh takes per setting, because your approved one was picked from a few takes. Face match is how much the face looks like you (we drop anything under 0.72).

Short answer: low (15 credits) keeps the same layout and idea as high (90), and the face still looks like you on 4 of 5 (0.75 to 0.85). It loses texture: on the giant chip, one low take lost the chip's ridges even at thumbnail size, and up close skin looks airbrushed. On a hard face (squinting, small in the frame) it can drift: one low BB take fell to 0.56. Medium (60) kept the chip ridges but shared low's smooth look elsewhere, so it is not a clear middle. Two takes per setting is a small sample: the BB scores vary more between takes than between settings. Earlier today a different test lost the face badly at low (0.79 at high, 0.25 at low) on a small face in an action scene, and a high pass over a low draft did NOT bring the face back. So: low is fine for drafts to find the idea; a low draft you like is a different picture from its high version. Low and medium prices are from Magnific's docs (high 90 was measured). Chips: same prompt, but the reference photos were rebuilt (the first run did not save which face photo it used).

Thumbnail	Approved	High (90)	Medium (60)	Low (15)
one chip bigger than his chest	0.79	0.76 / 0.74	0.77 / 0.81	0.78 / 0.75
one tiny chip, SCAM?	0.91	0.88 / 0.90	0.88 / 0.83	0.85 / 0.85
real vs 3D printed revolver	0.87	0.81, 1 refused	0.80, 1 refused	0.79 / 0.80
giant printed Buster Sword	0.83	0.80 / 0.78	0.79 / 0.77	0.79 / 0.79
one BB frozen at his face	0.76	0.67 / 0.75	0.65, 1 refused	0.56 / 0.70

Jump to: can you upscale a low one later?

Snacks: one chip bigger than his chest

The clearest loss. Low take 1 lost the chip's ridges and looks like a flat pale disc, visible even at thumbnail size; both high and both medium takes keep the ridges and salt. Low skin is smoother. Face match 0.74 to 0.81 everywhere.

**Your approved one (high)** face match 0.79

**Your approved one (high)** face match 0.79

**High (90 credits), take 1** face match 0.76

**High (90 credits), take 2** face match 0.74

**Medium (60 credits), take 1** face match 0.77

**Medium (60 credits), take 2** face match 0.81

**Low (15 credits), take 1** face match 0.78

**Low (15 credits), take 2** face match 0.75

Snacks: one tiny chip, SCAM?

Hardest to tell apart: all six look like the one you approved. The low chip, and one medium chip, is a smooth disc instead of a crinkly one. Face match 0.83 to 0.90.

Satisfying 3D Prints: real vs 3D printed revolver

Low is fine at thumbnail size. GPT refused 2 of the 6 (real-looking guns): one high, one medium, no low. Low was not refused more often.

Refused by GPT
content violation, nothing made

High (90 credits), take 2

Refused by GPT
content violation, nothing made

Medium (60 credits), take 2

Satisfying 3D Prints: giant printed Buster Sword

Very close at all three. Up close, the 3D-print layer lines on the gold guard are fainter at low (and on one medium take), and skin is smoother. Face match 0.77 to 0.80.

DIY Builds: one BB frozen at his face

The one where low hurts most. A squinting face that is small in the frame is hard at every setting (high 0.67 and 0.75, medium 0.65), and low take 1 drifted to 0.56 and reads as a different guy; low take 2 (0.70) actually beat high take 1. One medium take was refused.

Refused by GPT
content violation, nothing made

Medium (60 credits), take 1

Can you upscale a low one later?

Yes, you can, and it comes out sharper and cleaner, but it does not become the high version. This is the same upscaler and setting your approved finals used (Magnific Precision v1, 2x): it sharpens what is there and does not invent new texture. So the low chip stays flat and matte next to the glossy, crunchy high chip, the low BB pellets stay melted-looking blobs instead of round balls, and the sword blade stays softer. Face match did not move (each within 0.01: giant chip 0.78, tiny chip 0.85, sword 0.79, BB 0.70). Cost: low + upscale about 105 credits, high + upscale about 180 (upscale measured at 90). Part of each difference is the take, not the setting: every render is a fresh roll, so pose and light differ too.

Each picture: top row the face, bottom row the object, zoomed to the same size. Left: the low one blown up with no upscaler. Middle: the low one through the same Magnific upscaler your finals use. Right: your approved final (high, then upscaled). Tap to open full size.

Snacks: one chip bigger than his chest

Sharper, but the chip is still flat and pale; the high one is glossy with real crunch.

Snacks: one tiny chip, SCAM?

The closest of the four: after the upscale, face and logo hold up well next to the approved one.

Satisfying 3D Prints: giant printed Buster Sword

Cleaner edges and a faithful face, but the blade stays softer than the high one.

DIY Builds: one BB frozen at his face

The upscale sharpens the face but the BB pellets in the bottle stay blobby; the high one has distinct round balls.

Under each image: the face zoomed in, then a full-size detail crop from the middle of the frame (tap to open). Chips: original prompt, reference photos rebuilt (the first run did not save which face photo it used).