LaMa-Dilated-LiteRT

LaMa-Dilated — LiteRT (TFLite) GPU

⚠️ Archived / not maintained. This .tflite was produced by Qualcomm AI Hub's compiler, not by the Google-official LiteRT converter (litert-torch / ai-edge-torch), and LaMa is no longer part of my maintained conversion lineup. The artifact is kept for reference only (it was verified bit-identical to the qai_hub PyTorch reference, corr = 1.0). For maintained, reproducible LiteRT conversions see my other repos.

On-device LiteRT (.tflite) build of LaMa-Dilated large-mask image inpainting: paint a mask over an unwanted object and the model fills the region naturally.

This is the LaMa-Dilated variant (FFT blocks replaced with dilated convolutions for GPU friendliness), obtained from Qualcomm AI Hub. It is a pure CNN — all 361 ops are LiteRT GPU-native (no Flex ops, no CPU fallback on Qualcomm GPUs).

Files

I/O

• Inputs (NHWC, both [0, 1]):
  • image [1, 512, 512, 3] — RGB
  • mask [1, 512, 512, 1] — single channel, 1 = region to inpaint, 0 = keep
• Output:
  • painted_image [1, 512, 512, 3] — inpainted RGB

The model preserves the unmasked region exactly and only synthesizes inside the mask.

Usage (Android, LiteRT CompiledModel)

val model = CompiledModel.create(
    context.assets, "lama_dilated.tflite",
    CompiledModel.Options(Accelerator.GPU), null
)
val inputs = model.createInputBuffers()
val outputs = model.createOutputBuffers()
inputs[0].writeFloat(imageFloats)   // [1,512,512,3] RGB [0,1]
inputs[1].writeFloat(maskFloats)    // [1,512,512,1]  1=hole
model.run(inputs, outputs)
val painted = outputs[0].readFloat() // [1,512,512,3]

On-device note

The graph contains 3 TRANSPOSE_CONV ops. These run on Qualcomm Adreno GPUs (the hardware Qualcomm AI Hub optimized for). Some other GPU delegates may not accept TRANSPOSE_CONV; verify on your target device.

License & attribution

• License: Apache-2.0.
• Original model: advimman/lama (LaMa, © 2021 Samsung Research), Apache-2.0.
• Converted/optimized .tflite from Qualcomm AI Hub — lama_dilated.
• This repository only re-hosts the LiteRT .tflite for convenient on-device use; all credit to the original authors and Qualcomm AI Hub.

<!-- funnel:v1 -->

---

Want a different model on-device? Open a request — free, open weights only; the export and its measured numbers get published publicly.

<!-- /funnel:v1 -->