Qwen-Image 2.1 Turbo lokal auf dem Mac: Bilder Schritt für Schritt beim Entstehen zusehen

Posted on Oct 26 • Originally published at getoffgridai.co

You type a prompt and wait for an image. With most local image tools on a Mac you see a spinner until the last step, and only then find out the composition is wrong. In Off Grid AI Desktop (OGAD) the image now forms in the chat while it runs. Each sampling step draws a new rough frame, so you can see the layout settle and stop a bad image before it finishes. With Qwen-Image 2.1 Turbo, that is 8 steps from prompt to picture, all on your Mac.

If your Mac does not have room for Qwen-Image, the picks show SDXL models instead. Those still show live previews, but the Turbo defaults in this article apply only to Qwen-Image 2.1 Turbo.

This step is optional. Live previews work in both modes, but in-memory mode keeps the model loaded between images.

The app's own hint for this switch says in-memory cuts a typical cold start from about 45 seconds to about 7 seconds. That figure was measured on a small 512-pixel, 4-step image. Qwen-Image 2.1 loads a larger stack, so expect a longer first load, but the saving on the next image is the same idea: no reload.

On a Mac with 32 GB of memory or more, the in-memory image model now also stays loaded while you chat. A chat turn no longer pushes it out, so the next image starts without loading the Qwen-Image stack again. Below 32 GB, chat and image models still take turns, because the chat model plus the Qwen-Image stack needs about 16 GB together. A job that also frees the chat model, such as video, still frees the image model.

A concrete prompt makes the preview easier to judge:

A ceramic coffee cup on a wooden desk by a window, morning light, a small plant beside it, the word "OPEN" on a paper sign in the background.

While the model runs, a frame appears in the chat above a status line that counts the steps, for example Step 3 of 8, and a green bar that fills as steps finish. The frame updates at every step.

For Qwen-Image 2.1 the frames are shown in greyscale. They are a quick projection of the work in progress, which needs no extra model, so they show shape and layout, not final colour or detail. Look for the things you can judge early: where the subject sits, whether the framing is right, whether a sign or label is in the right place.

When the last step finishes, the status reads Finalizing image and the full-colour image replaces the preview. It is saved in the conversation with its size, steps, CFG and seed.

If the layout is wrong at step 2 or 3, click the red Stop button in the message box and change the prompt. With 8 steps, you learn early whether the image is going where you want.

Yes. OGAD 0.0.55 runs it on Apple silicon Macs with macOS 13 or later. Download it from Models, Image tab, and generate from chat with Generate image.

The app lists 32 GB of unified memory for the Full quality (Q8_0) size and 24 GB for Light (Q4_K). The model's published benchmarks put peak memory near 13.1 GB for the 8-bit build and 9.8 GB for the 4-bit build, and your chat model needs memory too.

For Qwen-Image 2.1 the app shows the step frames in greyscale, because they are a fast projection of the unfinished image. The final image is in full colour.

Turbo is the distilled version. It uses 8 steps and Guidance 1.0 where the full model uses 40 steps and Guidance 6, with the same text encoder and VAE. The full model is the Best pick when you want the most detail.

Yes, once the model files are on your Mac. Generation runs on your Mac with the local image engine, and the prompt and image stay on your machine.

No. Previews appear in both on-demand and in-memory mode. In-memory mode removes the reload between images, and on a Mac with 32 GB or more it keeps the image model loaded while you chat.