A reference photo says more than a long visual description.
OGAD (Off Grid AI Desktop) lets you use that photo as the starting image for local generation. You can explore a different mood, material or illustration style while giving the model a visual starting point. Download a compatible image model first; the editing workflow then runs on your computer without a cloud AI upload.
Download OGAD for Mac or Windows
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.
Use the reference for a clear purpose
Choose one thing the photo should contribute: the arrangement of objects, the broad composition or the starting colors. Then describe the change you want. A useful first task is turning a simple still-life photo into an illustration for a personal project.
For example, photograph a cup beside a notebook and a plant. The photo supplies the arrangement. Your prompt supplies the new treatment:
An editorial watercolor illustration of a cup beside a notebook and a small plant. Soft morning light, pale green and warm cream palette, simple background, no lettering.
The result is a new interpretation. It can move objects, alter details or invent parts of the scene. Do not depend on this route to preserve an exact product, face or logo.
Prepare a reference the model can use
Use a local image with a clear subject and enough surrounding space for the result you want. Crop out distractions before attaching it. Avoid combining several unrelated reference photos for the first attempt; this workflow uses one starting image for the edit.
If your photo is only in cloud storage, download it before your offline session. Keep a separate original so you can start another version without treating a generated result as the source of truth.
Make the first variation
- In Models, download and select a local image model with image-to-image support. The released SDXL Lightning and Juggernaut XL variants include this mode.
- Open Chat, then + → Generate image.
- Open Image options, choose + Init image, and select your local source image.
- Check that its filename appears beside Strength.
- Enter your prompt and generate one draft.
- Compare it with the source, then open the result and select Download to save it.
Choose a supported local SDXL image-to-image model for this route. In the checked release, Z-Image, Core ML and MLX image paths reject this type of reference-image edit. The ability to generate a new picture does not establish support for editing an existing one.
Local image generation is in the free desktop app. Model download size is separate from the memory needed while generating; use the app’s hardware guidance and begin with a modest output size.
Control how far the image changes
Strength controls how strongly the generation moves away from the starting image in this image-to-image route. Lower values retain more of its influence; higher values allow more change. It is not a guarantee that specific objects or text will remain untouched.
Try a moderate value first. Keep the prompt fixed and lower Strength if the arrangement changes too much. Increase it a little if the result stays too close to the original to be useful. Compare the results side by side.
Once you like the arrangement, change the mood in the prompt. For example, replace “soft morning light” with “cool overcast light.” Making one change at a time gives you a clearer view of what caused the difference.
Make another variation from the same photo
Return to the same conversation and choose + → Generate image > Image options. Check the filename beside Strength. Keep it only if it is the original photo you want. To change it, clear the selected image and use + Init image to choose the original again.
Keep the model, size, and Strength fixed, then change one part of the prompt. Save each useful result under a different filename. This compares new interpretations of the same source, rather than repeatedly feeding an altered result back into the model.
If you want to develop a generated image instead, select that saved output as the new starting image deliberately. Keep the original file so that you can return to it later. Reopening a conversation cannot restore a source file that you deleted or moved.
Use the result as a concept you can refine
This is useful for choosing a direction before detailed design work: a room mood study, a poster illustration or a still-life treatment. It gives you an image to respond to rather than asking you to imagine every word of a brief.
Check the details that matter for your project. Add exact text in your design tool. If a required object shape changes, keep the original reference visible and treat that output as unsuitable, even if the picture looks attractive.
This guide follows the local reference-image controls in OGAD 0.0.51. Use a complete compatible model and leave optional prompt enhancement off for the first run if you want to avoid another model dependency.
Try one reference and one change
Download OGAD, choose a photo you own, and make one deliberate variation. Keep the arrangement as your reference and change the visual treatment. You can explore the idea locally, with the original beside you.