Run and compare TinySD image generation results

TinySD Studio adapters

TinySD Studio uses an adapter to connect its shared prompt, seed, model import, and image display interface to a model runtime and the model package’s tokenization, denoising, and decoding requirements. When the application starts, it selects the registered model and chooses the matching adapter automatically.

The supplied TinySdImageGenerationAdapter owns the complete TinySD ExecuTorch workflow. It imports the package, initializes the tokenizer and model runner, generates the image, reports progress, and releases the runtime resources.

An adapter is application code compiled into the APK. Importing a model package doesn’t import executable application code or create another adapter. You’ll examine the adapter implementation on the next page. For the runtime API, see Run ExecuTorch on Android .

Image generation with TinySD Studio

You’ll run TinySD INT8 with ExecuTorch and XNNPACK. The application accepts a text prompt and a whole-number seed, then creates a 512 × 512 image entirely on the Android device.

The model package isn’t included in the APK. The supplied downloader retrieves the optimized ExecuTorch program, tokenizer, and scheduler data, then creates the uncompressed ZIP package expected by the adapter.

Run TinySD with ExecuTorch

Set the Arm model repository ID and run download_model.py from the cloned application directory:

    

        
        

export MODEL_ID="Arm/tiny-sd-int8-xnnpack-executorch-vivo-x300"

export MODEL_PACKAGE="$(.hf-venv/bin/python download_model.py \
  --repo-id "$MODEL_ID" \
  --print-path)"

printf 'Model package: %s\n' "$MODEL_PACKAGE"
  

    

    

        
        

$MODEL_ID = "Arm/tiny-sd-int8-xnnpack-executorch-vivo-x300"

$MODEL_PACKAGE = .\.hf-venv\Scripts\python.exe download_model.py `
  --repo-id "$MODEL_ID" `
  --print-path

Write-Output "Model package: $MODEL_PACKAGE" 
  

    

The downloader creates tinysd_vivo_executorch.zip, which is approximately 1 GB. It contains three files under the entry names required by the Android importer:

• huggingface/optimized.pte: the optimized multimethod ExecuTorch program
• huggingface/schedule_data.json: constants for the 25-step DPM-Solver++ denoising loop
• tokenizer/tokenizer.json: the CLIP vocabulary and byte-pair encoding rules

Copy the package to the Android Downloads directory:

    

        
        
adb push "$MODEL_PACKAGE" /sdcard/Download/tinysd_vivo_executorch.zip

    

In TinySD Studio:

1. Select Add or change model.
2. Open Downloads in the Android document picker.
3. Select tinysd_vivo_executorch.zip.
4. Keep the application open until it reports Model ready.

The importer checks the required ZIP entries and copies them to application-specific storage. You can delete the ZIP from Downloads after import. The imported files remain until you uninstall the application or clear its storage.

Enter the following prompt:

    

        
        
A watercolor lighthouse on a cliff at sunset

    

Leave the seed set to 42, then select Generate image.

TinySD Studio first runs the text encoder. It reports each of the 25 denoising steps, then runs the VAE decoder. Don’t close the application while generation is running. The first image can take several minutes, and timing varies with the Android device or emulator host.

When generation finishes, the application displays the image, seed, elapsed time, resolution, and runtime:

TinySD Studio after generating an image on the Android device

Select Save image. Android opens the document picker with a suggested .png filename. Choose a location such as Downloads, change the filename if needed, and select Save.

Compare another seed

Keep the prompt unchanged, replace the seed with another whole number, and select Generate image again.

Changing only the seed keeps the prompt embeddings the same but starts generation from different latent noise. The result should depict the same requested subject with a different composition or visual details. Reusing the same model, prompt, and seed reproduces the same variation.

This comparison isolates the effect of the seed. You can then change the prompt to see how different text embeddings guide generation.

Use another supported model

TinySD Studio currently has one registered and validated model package. Complete the guided workflow before changing the application or downloader.

The repository also contains an FP32 ExecuTorch program. It isn’t registered or validated with this application, and the downloader deliberately packages the optimized program only.

What you’ve accomplished and what’s next

You’ve downloaded and packaged the optimized TinySD model and imported it into TinySD Studio. You’ve also generated and saved an image locally, and compared the effect of another seed.

Next, you’ll learn how the application resolves the adapter and runs the TinySD image-generation pipeline.