[KongchangAI]
ModelERNIE-Image / 文心一格

ERNIE-Image Turbo

An image generation model in Baidu's ERNIE series that supports text-to-image generation. It is available in two forms: a base version and a distilled Turbo version. The model can be run locally.

Timeline (last 90 days)

Oct 16

On a Mac Studio (Apple M3 Ultra, 96 GB memory), in tests using mflux 0.19.1, 8-bit quantized weights, and 1600x896 resolution, ERNIE-Image Turbo averaged a score of 4.28 (out of 5) while FLUX.2 klein 4B averaged 3.07.

Unverified50%
Oct 16

FLUX.2 klein 4B's median time in the test was 22 seconds, roughly 3.6 times faster than ERNIE-Image Turbo's 80 seconds.

Unverified50%
Oct 16

ERNIE-Image Turbo received the highest score (alone or tied) on five of the six prompts.

Unverified50%
Oct 16

To judge whether a 32 GB Mac can run a given model, one should refer to macOS's peak memory footprint (39.0 GB for ERNIE-Image Turbo), not RSS (which reported only 5.1 GB in the test).

Unverified50%
Oct 16

In the counting prompt "five apples plus two pears," only ERNIE-Image Turbo's two seeds and the ERNIE-Image base version drew both the quantity and positions correctly (4.75 points); both FLUX.2 klein seeds drew only four apples.

Unverified50%
Oct 16

In the keyboard product image test, the judge deemed the key layouts of seven of the eight images impossible, and no model could draw a plausible keyboard.

Unverified50%

All Facts (6)

Source Articles