Can Ming Image run offline on Mac?

Yes. After the model download finishes, Ming generation runs locally in Radiant Canvas without an internet connection.

Is Ming 8-bit faster than 4-bit?

On M5 Max with 48 GB at 1024 × 1024, 8-bit took 27.1 seconds in Standard mode and 15.1 seconds in Fast mode. The 4-bit option took 27.2 seconds in Standard mode and 15.2 seconds in Fast mode. The 0.1-second difference between precisions does not establish a general speed advantage.

Can Ming make a transparent background?

Yes. Ask for one in your prompt. Ming can generate RGBA PNG images with transparency. Inspect the edges because faint background pixels can remain.

Can I use Ming to edit an existing photo?

Ming in Radiant Canvas is for text-to-image generation. For reference image editing, choose FLUX.2 Klein or Qwen Image Edit 2511 in the Edit tab.

Ming Image on Mac: Text & Transparent Images

What Ming does in Radiant Canvas

Ming turns a text prompt into an image in Create or a Studio Text to Image node. Try a square poster, an illustrated title, or a separate object you want to place in a design. The output can include transparency. Generated lettering is part of the image, so check every word before using it.

A local Mac app, with no separate setup

Radiant Canvas runs Ming on Apple silicon. Download the model in Model Manager, then generate without Python, terminal commands, or a separate server. Prompts and images stay on your Mac. Both Ming precisions are included in the free app; optional PRO tools vary by model.

Mac requirements and download sizes

You need an Apple silicon Mac running macOS 26.2 or later, with at least 16 GB of unified memory for either Ming option. Model Manager checks compatibility and available storage before downloading. More memory lets the app keep more of the model loaded.

Ming 4-bit: about 13.4 GB to download. It is the smaller option if storage is limited.
Ming 8-bit: about 25.1 GB to download. It retains more precision in the model weights, but that does not guarantee a better image.

These are download sizes, not estimates of peak memory use. Start with one precision, try your own prompts, and add the other if you want to compare results.

Make your first Ming image

Open Model Manager and search for Ming-Image-0.1-Design.
Choose 4-bit or 8-bit, check the requirements, and download it.
Open Create and select the installed Ming model.
Start at 1024 × 1024. Describe the subject, layout, colors, and any words you want in the image.
Generate and inspect the result. Adjust the prompt if the layout, wording, or transparent edges need work.

For a saved workflow, choose Ming in a Studio Text to Image node. Basic Studio text-to-image workflows are included free.

Write prompts for text and layout

Say what the image is for, where the main objects should go, and which words should appear. Put the exact wording in quotes. Short text makes it easier to inspect spelling and spacing. These examples are starting points, not guaranteed results.

A square poster for a garden workshop. Cream background, a green watering can in the lower half, and the words "GARDEN DAY" in large dark green letters at the top. No other text.
A square title illustration with the words "NIGHT WALK" in pale blue letters. Small stars around the title, a dark navy background, and wide margins.
A red ceramic mug viewed from the front, centered, with a transparent background. Keep the whole handle visible. No text.

Check transparent backgrounds before using them

Ming produces RGBA PNG images, which can contain transparent pixels. Ask for a transparent background in the prompt. The app preserves the transparency generated by the model; it does not automatically clean up faint background pixels.

Check the result against both a light and a dark background, especially around hair, soft shadows, and object edges. Some images may need cleanup in an image editor. A transparent image is still a single image, not a set of editable design layers.

1024 or 2048 output

Ming supports square output at 1024 × 1024 or 2048 × 2048. Start at 1024 while trying prompts. Larger output takes more work and can use more memory.

On Macs with less than 24 GB of unified memory, a 2048 request is generated at 1024 and then resized to 2048. It is not native 2048 generation on those Macs. Ming does not support portrait or landscape output in the app.

Ming 4-bit and 8-bit benchmarks on M5 Max

Measured at 1024 × 1024 on a MacBook Pro with M5 Max and 48 GB unified memory, with Standard mode using default settings (12 steps). In Standard mode, Ming 4-bit took 27.2 seconds and 8-bit took 27.1 seconds. In Fast mode, 4-bit took 15.2 seconds and 8-bit took 15.1 seconds. Fast mode requires PRO. The 0.1-second difference does not establish a speed advantage for either precision. Times vary with your Mac, prompt, and settings.

These results are generation timings, not reference editing timings. Choose between the two downloads based on storage and the images you get from your own prompts. The two precisions have nearly identical times within each mode in this test.

When to use another model

Ming in Radiant Canvas supports text-to-image generation only. It does not accept reference images and does not support Image Edit, Image to Image, or Inpaint nodes. Use FLUX.2 Klein for one-reference editing or Qwen Image Edit 2511 for up to three references.

Ming does not currently support LoRA or LoKr adapters, negative prompts, or previews during generation. PRO does not add those capabilities to Ming. The separate Ming Design-Layer model is not included in Radiant Canvas.

Use Ming Image on Mac for text and transparent backgrounds