Mac Local Image Generation Models Tested: ERNIE vs FLUX

Running local image generation models on a Mac Studio, ERNIE-Image Turbo leads with 4.28 out of 5, ahead of FLUX.2 klein 4B's 3.07. This article is based on blind-test data from six models, six prompts, and 48 images, offering directly reproducible selection advice and licensing reminders.

Test Design: Six Models, Six Prompt Sets, Blind Scoring

The tests were completed on a Mac Studio (Apple M3 Ultra, 96 GB memory), using mflux 0.19.1 with 8-bit quantized weights, all at 1600x896 resolution. Timing includes loading weights from the SSD, since that's what you actually wait for after typing the command. Scores came from two AI judges (Claude Opus and Claude Sonnet running independently), who rated images on four dimensions from 0-5 without knowing which model produced them, and the scores were averaged. ERNIE-Image Turbo and FLUX.2 klein each ran two random seeds, making it easy to observe how a single roll of the dice affects results.

The prompts cover common real-world needs: scene following, text spelling, photorealism, diagrams, composition and lighting, and counting. All models submitted results for seed 11, and judges received randomly named, metadata-stripped pure PNG files, unblinded only afterward.

mflux is an image generation toolkit built on Apple's MLX framework, specifically optimized for Apple Silicon's unified memory architecture. MLX is an array computing framework Apple developed for M-series chips, similar to PyTorch but directly calling Metal and unified memory, avoiding explicit CPU-GPU copies. mflux builds inference pipelines for diffusion models like FLUX and ERNIE on top of it, and provides command-line tools (such as mflux-generate-flux2). 8-bit quantization means model weights are compressed to 8-bit integers per parameter, roughly halving memory usage compared with 16-bit, but usually with slight quality loss; 4-bit compresses further, with more noticeable quality loss but possibly faster speed due to reduced memory bandwidth pressure.

Overall Ranking: ERNIE-Image Turbo Leads, FLUX.2 klein Fastest

In total scores, ERNIE-Image Turbo averaged 4.28 across 12 images, with 4.40 for the single seed 11 run. ERNIE-Image base scored 4.02, Qwen-Image-2512 scored 3.83, and FLUX.2 klein 4B only 3.07. ERNIE-Image Turbo took the top score (alone or tied) on five of the six prompts.

On speed, FLUX.2 klein's median time was 22 seconds, 3.6x faster than ERNIE-Image Turbo's 80 seconds. Boogu took 55 seconds, Z-Image Turbo 74.5 seconds. But the fastest model scored only 3.35, and slow models didn't buy higher image quality either — ERNIE-Image base took 813 seconds yet scored 0.38 lower than its distilled Turbo version.

Memory usage can't be judged by parameter count alone. ERNIE-Image Turbo's MLX peak memory was 22.5 GB, FLUX.2 klein 21.65 GB, while the smaller-parameter Z-Image Turbo (6B vs 8B) peaked at 34.48 GB. If you're judging whether a 32 GB Mac can run a model, refer to macOS's peak memory footprint (39.0 GB for ERNIE-Image Turbo), not RSS — RSS reported only 5.1 GB in tests, severely underestimating MLX's actual allocations.

The difference between RSS (Resident Set Size) and macOS peak memory footprint: RSS only counts the physical pages currently resident in main memory for a process, excluding allocations swapped out to swap or in GPU unified memory driven by Metal. MLX uses unified memory heavily on Apple Silicon, and those allocations may not show up in RSS. macOS peak memory footprint records the highest instantaneous usage of all physical memory (including GPU and neural engine) from process start to finish, so it's closer to actual pressure. When judging whether a 32 GB Mac can run a model, look at "peak footprint" rather than RSS, or you may severely underestimate real memory needs.

Scenario-by-Scenario Performance: Counting and Text Are the Dividers

Counting ability: only ERNIE and Qwen drew five apples correctly

In the "five apples plus two pears" prompt, only ERNIE-Image Turbo's two seeds and ERNIE-Image base drew both the correct number and position, all scoring 4.75. Qwen-Image-2512 also counted correctly, but the composition was crowded. FLUX.2 klein's two seeds both drew only four apples, Boogu drew them vertically as four, and Z-Image drew seven apples and three pears. Counting is the weakest area for these models, and if your prompt contains numbers, this test should weigh more than all the others.

Text spelling: FLUX.2 klein misspelled on both seeds

In the "OPEN SOURCE SALE" three-word banner test, FLUX.2 klein seed 11 spelled it "OPEN SPOOE", seed 42 spelled it "OOPEN SOURSCE SAL", and also generated extra gibberish signage. The other five models all spelled the banner correctly, with score differences concentrated in details: ERNIE-Image base had the cleanest image (4.38), ERNIE-Image Turbo seed 11 tied at 4.38 but drew the 2.99 price tag twice, and Qwen-Image-2512 spelled both lines of rounded type correctly (4.12).

Photorealism: no model could draw a plausible keyboard

Of the eight keyboard product images, judge A considered seven to have impossible key layouts — misplaced spacebars, jumbled key rows, fake letters on keycaps. ERNIE-Image Turbo seed 11 and Boogu tied for best at 3.75. FLUX.2 klein's two images earned the comment "best lighting, worst keyboard." If you need product shots of real objects, camera photography is still more reliable.

Diagrams: FLUX.2 klein's only top score

In the PLAN, BUILD, SHIP three-box flowchart test, FLUX.2 klein seed 11 won with 4.75, with colors matching exactly as required, only the rocket icon slightly blurry. But its seed 42 version scored only 4.00, with the wrench icon melting. All eight images spelled the three words correctly, showing short labels aren't hard for any model.

Composition and lighting: the lighthouse in the storm was a giveaway

This group's scores clustered between 3.88 and 4.75. Both ERNIE versions scored 4.75. FLUX.2 klein's two seeds both drew the lighthouse^Ad too small to carry the frame (4.00 and 3.88).

Licensing Traps: 4B and 9B Fall Under Different Agreements

All six tested models are Apache-2.0 (three of their Hugging Face repos have no LICENSE file, only a license field on the card). The real trap lies in models not tested.

FLUX.2 klein 4B is Apache-2.0 and commercially usable. But its sibling versions FLUX.2 klein 9B and FLUX.2 dev use the FLUX non-commercial license, whose terms are internally contradictory: section 2.d allows outputs for any purpose (including commercial), while section 4.a prohibits using data produced by the model for commercial or production purposes. For scenarios like a company blog, it's a no-go without separate authorization from Black Forest Labs.

Qwen's two image models are easily confused. Qwen-Image-2512 is Apache-2.0 and commercially usable; Qwen-Image 2.1 uses the Qwen research license, explicitly stating "FOR NON-COMMERCIAL PURPOSES ONLY", with commercial use requiring separate contact. Ideogram 4's free weights come with a non-commercial agreement, but its official FAQ says outputs can be used commercially, a conflict. Krea 2 permits commercial use only when company annual revenue is under $1 million, and FLUX 3 Image Klein's threshold is $5 million.

Some popular articles get this wrong. AI Weekly said Krea 2 is "free for teams under 50 seats," but the license has no seat limit at all, only a revenue line. Pixazo said FLUX.2 klein requires commercial authorization, which holds for 9B but not 4B. Another guide labeled 9B as Apache-2.0, inconsistent with the non-commercial license in its repo.

Apache-2.0 is a permissive open-source license allowing free commercial use, modification, and redistribution, requiring only retention of copyright notices and license text, and not requiring derivative works to use the same license. Non-commercial licenses (such as the FLUX Non-Commercial License) prohibit using the model or its outputs for commercial purposes, usually also restricting production use. Different parameter versions in the same model family may use different licenses, possibly due to training data source differences, partner requirements, or business strategy. For example, FLUX.2 klein 4B uses Apache-2.0, while its 9B version uses the FLUX non-commercial license, with different LICENSE files on Hugging Face. Before downloading a model, always check the LICENSE file in the corresponding repo, not just the model card summary.

Practical Advice: How to Run It on Your Mac

Install mflux and pin the tested version:

uv tool install mflux==0.19.1

Point the Hugging Face cache to a large-capacity disk — ERNIE-Image Turbo's cache is about 22 GB, Qwen-Image-2512 about 54 GB:

export HF_HOME=/Volumes/your-big-disk/huggingface

Generate a banner test image with ERNIE-Image Turbo:

mflux-generate-ernie-image-turbo --steps 8 -q 8 --seed 11 --width 1600 --height 896 --metadata \
  --prompt "Street photograph of a small corner shop front in daylight. A large red banner above the window reads OPEN SOURCE SALE in white block capitals, and a yellow price tag in the window reads 2.99. No people, no other text." \
  --output sale-ernie.png

Then run FLUX.2 klein 4B with the same prompt for comparison:

mflux-generate-flux2 --model flux2-klein-4b --steps 6 -q 8 --seed 11 --width 1600 --height 896 --metadata \
  --prompt "Street photograph of a small corner shop front in daylight. A large red banner above the window reads OPEN SOURCE SALE in white block capitals, and a yellow price tag in the window reads 2.99. No people, no other text." \
  --output sale-flux.png

On a 96 GB Mac, ERNIE-Image Turbo 8-bit at 1600x896 peaks at 22.5 GB MLX memory and 39.0 GB macOS peak footprint. 4-bit at 1024x1024 can bring the footprint down to 27 GB with no loss in counting ability. FLUX.2 klein 4-bit uses the same memory as 8-bit, for reasons that remain unclear.

Conclusion: The Speed-Quality Tradeoff

ERNIE-Image Turbo is the most balanced choice for Mac local image generation between quality and speed. If you need quick drafts, FLUX.2 klein 4B's 22 seconds is still valuable. But don't assume every version is commercially usable just because of the word "open source" — the same family's 4B and 9B may fall under different licenses. Before publishing anything, read the license file for the exact checkpoint you downloaded.

A single seed and a single prompt can't form a reliable ranking. If your work is mainly product images, it's worth rerunning with your own ten prompt sets.

One-line Summary

Mac local image generation blind test: ERNIE-Image Turbo leads in quality, FLUX.2 klein is fastest, and licensing traps deserve caution.

Paragraph Summary

This article blind-tests six local image generation models on a Mac Studio (M3 Ultra, 96GB), using mflux 0.19.1, 8-bit quantization, 1600x896 resolution, six prompt sets totaling 48 images, scored by two AI judges on four dimensions. Results show ERNIE-Image Turbo leads with 4.28, while FLUX.2 klein 4B scores only 3.07 but is fastest (median 22 seconds). Counting and text spelling are the dividers; only ERNIE and Qwen drew five apples correctly, and FLUX.2 klein misspelled the banner on both seeds. On memory, parameter count doesn't determine peak usage — refer to macOS peak footprint. On licensing, FLUX.2 klein 4B is commercially usable, while 9B and FLUX.2 dev use non-commercial licenses; Qwen-Image-2512 is commercially usable, and Qwen-Image 2.1 is research-only. The article ends with reproducible commands and selection advice.