You have a photo and an idea for a different version: a new background, a different material or another visual style. You want to explore that idea without uploading the original.
OGAD (Off Grid AI Desktop) can run Qwen-Image 2.1 locally for image generation and reference-image editing. This guide uses released beta 0.0.52-beta.103 with manual model setup. Download the app and four model files first. Generation can then run without internet when you use the local image model.
Download OGAD beta 0.0.52-beta.103
What would you like to do with Off Grid AI?
Have a feature or use case you would like us to support? Tell us what you want to do and which device you use.
Write to support@offgridmobileai.co, join our Slack community, or talk to us on Reddit.
What do you need?
- A Mac running macOS 13 or later. Plan for 24 GB of RAM or more for this model stack.
- About 11.06 GB of storage for the four model files, plus space for OGAD and your generated images.
- Internet for the initial downloads.
Download size is not working memory. The 24 GB figure is setup guidance based on the app's memory check, not a speed or reliability guarantee. Close other large apps before generation.
Image generation is part of OGAD's core features. The Qwen-Image 2.1 model has its own research license; check its terms for your intended use.
How do you install the complete model?
Install the Mac .dmg from the linked beta release. Open OGAD once, then quit it.
This beta can run Qwen-Image 2.1, but its model catalog does not offer the complete download. Download these exact four files yourself. Keep their names unchanged.
| File role | Download | Approximate size |
|---|---|---|
| Image model | qwen_image_2.1-Q4_K.gguf |
4.20 GB |
| Text encoder | Qwen3VL-8B-Instruct-Q4_K_M.gguf |
5.03 GB |
| Vision projector for editing | mmproj-Qwen3VL-8B-Instruct-F16.gguf |
1.16 GB |
| Qwen-Image 2.1 VAE | qwen_image_2.1_vae_bf16.safetensors |
676 MB |
In Finder, choose Go → Go to Folder and paste:
~/Library/Application Support/Off Grid AI Desktop/models
Create the models folder if it does not exist. Put all four files directly inside that folder. Do not leave the VAE in a separate vae folder. These are the default paths; a custom OGAD data-directory override uses its own models folder.
The image model needs its text encoder and VAE. Reference-image editing also needs the vision projector. The older Qwen-Image VAE is not a substitute for the 2.1 VAE. The image runner's Qwen-Image 2.1 guide documents that model combination.
How do you generate your first image?
Restart OGAD after the files finish downloading. Use a local image configuration for this guide; a selected remote image server processes images elsewhere.
- Open Chat and use + → Generate image.
- Open Image options. If you have multiple local image models, choose qwen_image_2.1 in Model. With only one local image model, OGAD selects it without showing that dropdown.
- Set Size to 1024, Steps to 40 and Guidance to 6. Leave optional prompt enhancement off for this first image.
- Enter a prompt and send it.
Try a concrete visual brief:
A square editorial illustration of a small reading room, warm afternoon light, green armchair, wooden shelves, simple shapes, no text.
When the result appears, use Download to save it. Change one part of the prompt for the next version, such as the lighting or the color palette. That makes the comparison more useful.
How do you edit an existing image?
Keep Qwen-Image 2.1 selected. In Image options, choose + Init image and select your source image. Describe the change you want:
Replace the background with a quiet garden at sunset. Keep the main chair as the subject.
Send the request and compare the result with the original. This is generative editing: it can change details you wanted to keep. Check faces, lettering and product shapes before you use the result.
Use the prompt to direct the edit. The generic Strength control does not change Qwen-Image 2.1's edit request in this beta.
What if the model does not run?
| Problem | What to check |
|---|---|
| Qwen-Image 2.1 does not appear | All files finished downloading, their names are unchanged, and the primary GGUF file is directly in models. Restart OGAD. |
| Missing encoder or VAE error | Check the exact companion filenames and the folder above. |
| Generation works but editing fails | Check that mmproj-Qwen3VL-8B-Instruct-F16.gguf is present. |
| Not enough memory | The four files load as a stack. Use a machine with more memory or a smaller image model. |
The manual setup takes some preparation. After that, you can keep a prompt, its reference image and the generated result in one local chat.
Install the OGAD beta, add the four files and make one image from a brief you already want to explore.
Top comments (0)