VOSR_1.4B_Mobile

VOSR 1.4B Mobile

AllGPTORG/VOSR_1.4B_Mobile is a mobile deployment package of the VOSR-1.4B one-step image super-resolution model, quantized and compiled as Qualcomm QNN DLC graphs for the Snapdragon 8 Gen 3 NPU.

The package is designed for high-quality image restoration and super-resolution, including text-rich images. It preserves the original one-step VOSR pipeline while splitting the network into smaller graphs suitable for mobile integration.

This repository contains QNN deployment artifacts. It is not a Transformers or Diffusers checkpoint and cannot be loaded with from_pretrained().

Model summary

Files

The graphs must be executed in the order shown below.

manifest.json contains the same graph order, precision assignment, and exact byte size for programmatic use.

Ready-to-run SM8650 EPContext package

sd8g3/qnn-context/ contains eleven pre-linked QNN context binaries for the Snapdragon 8 Gen 3 HTP. They were linked with QAIRT 2.45.0.260326154327, target DSP v75 / SoC model 57, and use O1 graph finalization to stay inside the device's 8 MiB VTCM budget.

The DiT is split into six contiguous stages: 00_06, 06_12, 12_18, 18_24, 24_30, and 30_36. RMSNorm pointwise multiplications are divided along the token axis before linking; a CPU reference comparison against the unmodified QDQ block measured 72.9 dB PSNR.

The complete context package is 1,140,031,488 bytes (1.062 GiB). The small EPContext ONNX wrappers and their exact SHA-256, byte-size, tensor, graph-order, and immutable Hub-revision contracts are stored in sd8g3/vosr_runtime_manifest.json.

The Qwen VAE decoder is compiled directly from the original FP16 ONNX graph. The earlier W8A16 QDQ decoder removed nearly all spatial detail. The FP16 SM8650 context matches the CPU ONNX decoder at 59.77 dB PSNR with a maximum per-channel error of one RGB level.

Raspberry Pi 5 ONNX Runtime package

pi/w8a16/ contains the CPU-oriented VOSR-1.4B runtime used by AllCamera Cloud AI. Its DINO and DiT MatMul weights use ONNX Runtime MatMulNBits INT8 quantization while activations remain FP16. VAE, prepare, and final graphs stay FP16. The package is 1.663 GiB, down from 3.116 GiB for the split FP16 graphs, and is pinned by the immutable Hub tag pi-w8a16-v1.

The graphs are deliberately executed one at a time so an 8 GB Raspberry Pi 5 does not hold the complete 1.4B pipeline in memory. The largest measured local FP16 graph peak was about 1.84 GB RSS; its quantized counterpart used about 557 MB. ONNX Runtime 1.28 provides an ARM64 wheel with the required MatMulNBits kernels.

Quantization was compared against the complete FP16 split-graph reference on a real zoom patch: 49.34 dB PSNR, 0.99986 RGB correlation, and matching edge energy. A full 40x 2 MP Cloud job passed geometry, color, border, noise, and detail gates after transferring the model reconstruction residual over the denoised OEM frame. This preserves real captured edges while still adding VOSR structure.

Tensor interface

All image and latent tensors use NCHW layout.

Integration outline

Use the Qualcomm AI Engine Direct SDK / QNN runtime to load and execute the DLCs. The host application is responsible for preprocessing, graph orchestration, random noise generation, the one-step latent update, tiling, and image postprocessing.

1. Resize the low-resolution image to the requested output resolution using bicubic interpolation, split it into 512 x 512 tiles if needed, convert RGB values to FP16 NCHW, and normalize them to [-1, 1].
2. Run the DINO graph and Qwen VAE encoder on the same image tile. Supply seeded normal noise as posterior_noise if reproducible output is required.
3. Create an FP16 normal-noise latent z with shape [1,16,64,64], concatenate lq_latent and z along the channel axis, and use the result as latent_pair.
4. For the one-step schedule, run DiT prepare with timestep = [1.0] and next_timestep = [0.0].
5. Pass hidden sequentially through all four DiT block-stage DLCs, or through all six sd8g3/qnn-context DiT stages when using the EPContext package. Reuse block_conditioning and projected_dino for every stage.
6. Run DiT final and apply the one-step update z = z - velocity.
7. Decode the updated latent with the Qwen VAE decoder, clamp the output to [-1, 1], convert it back to RGB, and blend overlapping tiles when tiling.

For 4x super-resolution, bicubic-upscale the source to the final target resolution before creating the model tiles. The network then restores detail at that target resolution.

Validation status

• The DINO, VAE encoder, DiT prepare/final, and DiT blocks use their documented mobile quantization. The VAE decoder deliberately remains FP16.
• All runtime graphs were successfully compiled and linked for SM8650.
• The Qwen VAE encoder and DiT prepare graphs were profiled on the Snapdragon 8 Gen 3 NPU.
• All eleven split EPContext graphs were linked and inspected as DSP v75 / SoC model 57 contexts. The runtime disables CPU execution-provider fallback.
• The complete on-device chain preserves generated detail: on the S24 Ultra regression patch, edge energy rose from 1.54 for the bicubic input to 15.19 after VOSR.
• The 00_06 six-block context was profiled successfully on the Galaxy S24 target: 382.9 ms estimated warm inference, 230.9 ms warm load, and 119.5 MB estimated peak memory. Every reported operator, including the VTCM chunks, executed on the NPU.
• A monolithic context is intentionally not included. Splitting the DiT into six VTCM-safe contexts avoids the cloud linker's memory explosion and allows the application to release each session before opening the next one.

Performance, memory use, and image quality depend on the QNN SDK version, device firmware, thermal state, tiling implementation, and host-side orchestration. Test on the exact target device before shipping a production application.

Intended use

This package is intended for research and mobile application development involving:

• photo and screenshot restoration;
• text-rich image enhancement;
• single-image super-resolution;
• Snapdragon NPU deployment experiments.

It is not intended for forensic reconstruction or for recovering information that is not present in the source image. Generative restoration can introduce plausible but incorrect details.

Attribution

This is a quantized mobile derivative of VOSR. The original architecture, training, and checkpoints were created by the VOSR authors. See the official project and paper for full details.

Citation

If you use this model, please cite the original VOSR work:

@inproceedings{wu2026vosr,
  title     = {VOSR: A Vision-Only Generative Model for Image Super-Resolution},
  author    = {Wu, Rongyuan and Sun, Lingchen and Zhang, Zhengqiang and Kong, Xiangtao and Zhao, Jixin and Wang, Shihao and Zhang, Lei},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year      = {2026}
}

License

Released under the Apache License 2.0, following the upstream VOSR repository. Users are responsible for reviewing and complying with the licenses and terms of all upstream components and the Qualcomm QNN SDK/runtime used for deployment.