Best Local Image Generation Model on Mac: ERNIE vs FLUX

Description and use

On a Mac Studio the author benchmarked six local image generation models. ERNIE‑Image Turbo delivered the best quality and accurate counting, while FLUX.2 klein 4B was fastest but made spelling errors. ERNIE‑Image Turbo was chosen as the default.

Problem

Choosing the best local image generation model for a Mac Studio with limited GPU memory and commercial licensing constraints.

Outcome

ERNIE‑Image Turbo produced higher quality images (average score 4.28) and correctly counted objects and spelled text, while FLUX.2 klein 4B was fastest (22 s median) but missed spelling and had licensing restrictions for commercial use.

Workflow

• 1. Install mflux 0.19.1 and 8‑bit quantised weights on Mac Studio.
• 2. Generate 48 images (six models × six prompts) at 1600×896 pixels, recording median seconds, steps, peak MLX memory, and macOS footprint.
• 3. Run two blind judges (Claude Opus and Claude Sonnet) to score each image on adherence, text accuracy, artefacts, and looks.
• 4. Compare scores, memory usage, and licensing information for each model.
• 5. Select ERNIE‑Image Turbo as default for future cartoons.
• 6. Document findings and update workflow.
• 7. Monitor performance and licensing compliance.
• 8. Publish results on dev.to.

Evidence

ERNIE-Image Turbo scored highest. Over its 12 images it averaged 4.28, and on seed 11 alone 4.40, against 4.02 for ERNIE-Image base, 3.83 for Qwen-Image-2512 and 3.08 for FLUX.2 klein.

Source

FLUX.2 klein is the fastest: a median of 22 seconds, 3.6 times quicker than ERNIE-Image Turbo's 80.

Source

ERNIE counted right. Asked for exactly five apples beside a bowl with two pears, ERNIE-Image Turbo drew five apples and two pears on seed 11 and on seed 42.

Source

FLUX.2 klein misspelled the three-word banner on both seeds: OPEN SPOOE on seed 11 and OOPEN SOURSCE on seed 42.

Source