Everything measured on one Mac Studio, the mothership
Image and comic generation
Memory
36GB
Chip
M5 Max
CPU
18 cores
GPU
32 cores
The mothership: Mac Studio (model MHL64J/A), macOS 27.0. Every number on this site was measured on this one Mac. Whether an AI runs at all comes down mostly to how much memory you have.
Not just text: AI that makes images can also run inside your own Mac. Nothing is sent over the internet, and there is no charge no matter how many you make. On this page, popular image AI models are given the same prompts on the mothership Mac Studio, and the time they take and the results are compared.
Selection guide
Models by task
Start with
FLUX.2 [klein] 4B
distilled / BF16 or app-supported quantization
Starting memory estimate
24 GB
For trying short tasks one at a time
A planning budget starting from the official ~13GB VRAM figure, with room for the text encoder, app and OS. Try one image at around 1024px. Format and app change actual usage.
Tasks to try
Prototype article illustrations, backgrounds and reference-guided edits. Generate one image at a time to refine composition.
Keep under review
Check exact Japanese text, logos and identity consistency separately. Adding text afterward may be more dependable.
Memory figures are editorial budgets including room for the OS and apps, not verified minimums. Text models assume 4-bit weights, 4K–8K input tokens and one model at a time. Long documents, images and concurrent apps need more. A3B does not mean the entire model fits in 3GB.
Assumes an Apple Silicon Mac and a compatible app: LM Studio or similar for text, Whisper apps for audio, and model-specific image/video tools. RAM alone does not determine speed. Start with a short input on your current Mac.
User feedback overview
What users value
Fast iteration with few steps receives positive feedback.
Recurring difficulties
Fine detail and realism can require tuning; increasing steps alone is not a guaranteed improvement.
A qualitative synthesis of public discussions, not a representative survey. Reports include PCs and cloud setups, so they do not establish Mac speed or memory requirements. No individual profiles or anecdotes are reproduced.
Our tests and conditions
Candidates to test
On hold or being checked again A failed run is recorded; no passing result is available · Updated
On 2026-11-14, FLUX.2 klein 4B (Q8) on a Mac Studio (M5 Max, 36GB) was asked to draw a four-panel comic from a fixed script. It produced four images in about 73 seconds, but even when asked for one panel at a time, each image came out split into four panels, the characters and scenes did not stay consistent, and panel 3 contained speech-bubble lettering, so it did not pass. No manual fixes were made. Attempts on 11/8 and 11/9 also did not pass. Next, methods other than the prompt, such as reference images, will be tried. The amount of memory used has not been measured yet.
MOSS-Image-SFT · MOSS-Image Base
Test tasks and measures
Compare Japanese lettering, panel layout, character consistency across panels and time to finish including manual lettering.
What we know and what remains open
Still-image times and examples are already available. Mac operation and finished work for the comic candidates remain untested.
Mac Studio No. 1 (36 GB); FLUX.2-klein-4B Q8 (digest 333efdfe…); mflux-generate 0.20.0, comic-visual-single-panel-v2. Input SHA-256: 889115c9a066b62dfa13d44f663c04b91212fc1a17cfc2c126eb0caffa636ab7. fixture_commit f7538ccf53b155a9019440d2a4df970c7183dee8.
Queue 20261109T032706-mac-llm-bench-84a1. Generated four PNGs in 71.5 seconds. Even with tightened single-panel prompt constraints, each PNG remained a 2x2 multi-panel comic; p2 had mismatched scene and characters, and p3 contained illegible glyphs, failing character consistency and no-text requirements. Independent assessment scored passed=false. No manual edits. Peak memory is null due to uninstrumented backend. This is not a passing result or proof for the full use case.
Cycle 1 (2026-11-08, same model and input, 81.7 s, job 20261108T055606-mac-llm-bench-340d, independent check failed) remains as a separate attempt.
Because tightening prompt constraints in cycle 2 still failed to produce single panels or eliminate glyphs, the next cycle should test under further refined single-panel and glyph-free constraints reflecting the prompt-following limits of the model.
Queue 20261114T194436-mac-llm-bench-836a. Generated four PNGs in 73.0 seconds. Even with per-panel generation, each PNG remained a 2x2 multi-panel comic. Independent review failed single-panel, character/scene consistency, and the no-text rule on source PNGs (p3 had speech-bubble lettering and glyphs). Panel order and fixed Japanese dialogue (HTML overlay) matched. Independent assessment scored passed=false. No manual edits. Peak memory is null due to uninstrumented backend. This is not a passing result or proof for the full use case.
Cycle 1 (2026-11-08, 81.7 s, job 20261108T055606-mac-llm-bench-340d) and cycle 2 (2026-11-09, 71.5 s, job 20261109T032706-mac-llm-bench-84a1) failures remain as separate attempts. Both used the same fixed input and model family and failed.
Because cycle 3 still produced 2x2 grids and glyph contamination under per-panel generation, the next cycle should try non-prompt controls such as reference images or post-generation single-panel enforcement, and also instrument peak memory and compare another model. Prior failures and the legacy measurement pause remain in place.
Published measurements
Results published Mac Studio · Apple M5 Max · 36 GB
Z-Image Turbo · Per image after the first 19.7 s · Peak memory 18.51 GB · 2026-09-24
FLUX.2 [klein] 4B · Per image after the first 5 s · Peak memory 16.67 GB · 2026-09-24
ERNIE-Image Turbo · Per image after the first 22.8 s · Peak memory 21.63 GB · 2026-09-24
Qwen-Image · Per image after the first 127 s · Peak memory 37.3 GB · 2026-09-24
Z-Image Turbo by Alibaba (Tongyi-MAI): about 19.7 s per image from the second one, about 22.1 s for the first, 18.5GB of memory, Japanese text not written Guide: a Mac with 32GB of memory
FLUX.2 [klein] 4B by Black Forest Labs: about 5 s per image from the second one, about 6.5 s for the first, 16.7GB of memory, Japanese text not written Guide: a Mac with 32GB of memory
ERNIE-Image Turbo by Baidu: about 22.8 s per image from the second one, about 26.6 s for the first, 21.6GB of memory, Japanese text written Guide: a Mac with 36GB of memory
Qwen-Image by Alibaba (Qwen): about 127 s per image from the second one, about 133.4 s for the first, 37.3GB of memory, Japanese text not written Guide: heavy even on a 48GB Mac, with other apps closed
Compare on the same prompt
Click an image to see it larger. Every image was made by AI and differs from real places and things.
To try it on your own Mac (Draw Things)
Install Draw Things. Search the Mac App Store for “Draw Things” and install it (free). It works with buttons and text boxes, so you don't need Terminal.
Pick a model. Choose one from the app's model list, for example “FLUX.2 [klein] 4B” or “Z-Image Turbo”. The first time, the model downloads (from a few GB to 10GB or more).
Describe the image and generate. Write a description of the image you want in the text box (the prompt) and press the generate button. Most models follow descriptions in English more closely.
The measurements on this page use a different tool that runs the same models (mflux). In Draw Things, speed and results can vary with the settings. Draw Things can also generate in the cloud; to keep everything on your Mac, don't use that feature and download the model instead.
How it was measured
Tool: each model runs under the same conditions in mflux, an image generation tool for the Mac.
First image: the time from loading the model until the first image is done. A guide to how long you wait the first time you try it.
From the second: the time per image when making several in a row.
Memory: the most memory used while making images (including the image processing part). The amount of used memory kept around for reuse is capped at 2GB for every model.
Japanese text: for the sign prompt, whether “こんにちは” (hello) was drawn is judged automatically by reading the image with the Mac's built-in text recognition (Vision). No person checks it, so garbled characters can be misread or misjudged.
Quantization: each model is quantized (made lighter) ahead of time and saved, and that saved copy is loaded. This is the same way you use the lighter model versions you can download in apps.
Images are 1024×1024, made with the same random number (seed) for every prompt.
Only models under a license that lets their images be published freely (Apache 2.0) are shown.
Release date: the day the model appeared on Hugging Face (when its repository was created). It can be a little earlier than the announced date.