You do not need a final-size image to decide whether an idea works. OGAD (Off Grid AI Desktop) lets you generate local image drafts, adjust the settings, and inspect progress as the picture takes shape. Start with a smaller draft and a suitable model, then spend more processing time on the direction you want to keep.
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.
That approach is useful for blog illustrations, character ideas, and presentation graphics. You can reject a poor composition before waiting for several large versions. The model runs on your computer after download, with no required cloud generation request.
What has the largest effect on waiting time?
The model, output size, number of generation steps, and your computer’s resources all matter. OGAD exposes size and step controls so you can compare a draft setup with a more detailed result. Use the selected model’s defaults as your starting point.
| Choice | Why it matters |
|---|---|
| Model | Different models require different amounts of work and memory |
| Image size | More pixels can require more generation work and memory |
| Steps | More steps add work; too few can damage the result |
| Prompt enhancement | Adds a text-model stage before generation |
| Other active work | Competes for the memory and processing the image job needs |
There is no single setting that is fastest and best for every model. A few-step model and a full model have different requirements. Compare useful results, not only the smallest number in the Steps field.
Make a draft before making the final image
Use a smaller supported size to test the subject, composition, and mood. In OGAD’s image settings, Size offers square dimensions including 512 × 512 and 1024 × 1024. Begin with a draft size the model supports, then inspect whether the idea deserves another pass.
For a blog illustration, the first question might be whether the main subject leaves room for a heading. Fine texture is less important at that stage.
Try this prompt:
A small green desk lamp on the right side of a wooden desk, open dark background on the left, warm evening light, simple illustration, no lettering.
Generate a draft and check the placement. If the lamp is centered, fix that problem in the prompt before requesting the larger version.
When you change the size, the composition can also change. Generating a larger image is a new generation, not a promise to enlarge the exact same pixels. Check the new result before using it.
Match the step count to the model
OGAD supplies defaults for the selected image model. Keep those for your first comparison. If you later reduce Steps, make a small change and check whether the result still has usable shapes, detail, and color.
A lower number can reduce work, but it can also produce an image you must discard. That is not a useful saving. Distilled or fast models are designed around different generation behavior from full models; do not copy one model’s settings to another without checking the output.
Change one setting at a time. Keep the prompt, model, size, and seed fixed when you compare step counts on the same setup. This makes it easier to see whether the change served your goal.
The controls described here are in OGAD 0.0.51. The useful settings depend on your model and computer.
Skip a rewrite when your prompt is already finished
The Enhance option can ask the chat model to add visual detail before image generation. That is useful for a rough request. If you have already written a complete prompt, turn it off to avoid that extra rewrite stage.
This also keeps a prompt comparison more controlled. You know that the image model receives the wording you wrote, without a new text-model rewrite on every attempt.
If you need help writing the prompt, use the assistant once, review the result, and reuse the finished text. You can then spend the next generations testing the visual choices rather than repeatedly revising the same brief.
What does the live preview tell you?
On supported generation paths, OGAD can show a rough preview while the image is being made. Progress stages also help you distinguish prompt enhancement, generation, and decoding. A job can still be working after the main generation steps finish.
Use the preview to see broad composition and whether the image is forming. Do not judge final detail from an early rough preview. Some backends do not provide the same preview path, so absence of a picture during generation is not by itself a failure.
If you decide to cancel, use the app’s stop control and wait for the job to stop before starting a replacement. Repeated requests can make it harder to tell which job is active.
Getting started with a useful comparison
Pick one prompt and make a baseline before changing the settings. Your goal is to find a setup that produces an acceptable draft sooner on your computer.
- Download and activate a local image model in Models.
- Open the image settings and keep the model’s default steps and guidance.
- Use a supported draft size and generate one image.
- Record the settings and whether the image is useful.
- Change one setting, generate again, and compare.
If the app shows generation time with the result, use it alongside your quality check. Include model loading in your expectations: the first request after a model change can differ from a later one.
Do not compare different prompts, models, sizes, and step counts all at once. You will not know which change caused the improvement or the problem.
What if generation is still slow or fails?
Check whether the computer has enough available memory for the chosen model and output size. Close demanding apps, stop other model work you do not need, and try a smaller supported image size. A model that cannot fit is a different problem from a slow but healthy generation.
If lower settings produce poor output, return to the defaults. You may get a better workflow by using another model for drafts than by pushing a detailed model below its useful settings.
Keep the selected route local if the reason you chose OGAD is to process images on your hardware. A remote server may have different speed and privacy trade-offs.
Spend the wait on the image you want
Download OGAD and test one prompt at a practical draft size. Choose the composition first, then generate the larger version you intend to use. A small, controlled comparison is more useful than guessing at the lowest settings.