A 9-channel SD1.5 inpainting UNet compiled for Qualcomm's Hexagon NPU, plus its own CLIP and VAE. Built for DreamUI; runs fully on-device, no network at inference time.

Converted from 5w4n/chilloutmix-inpainting.

The upstream checkpoint states no licence. 5w4n/chilloutmix-inpainting carries an empty model card with no license field, and ChilloutMix is a community merge whose provenance and terms are not clearly documented. These artefacts are redistributed on that basis, with attribution to the upstream repo, and no licence claim is made over the underlying weights. If you are the rights holder and want this removed, open an issue and it will be taken down.

Most SD1.5 derivatives are CreativeML OpenRAIL-M; that is an assumption here, not a statement of fact. Satisfy yourself before using these commercially.

The conversion pipeline derives from Local Dream (CC BY-NC 4.0), so treat the tooling lineage as non-commercial.

Each contains 7 flat files: unet.bin, clip_v2.mnn, pos_emb.bin, token_emb.bin, tokenizer.json, vae_encoder.bin, vae_decoder.bin.

⚠ Byte lengths are exact and verified by the downloader — do not re-compress or re-upload under the same name.

_min runs about 2.4× slower (9569 ms vs 4055 ms at 512x512 / 20 steps, measured on an S25 Ultra) but loads on far more hardware.

⚠ Both are QAIRT 2.49 builds, which stamp an fp16 execution requirement into the context binary. A device lacking that feature rejects the model at load with Request feature fp16 enable with value 1 unsupported, regardless of tier — _min lowers the arch and VTCM requirements, not the fp16 one.

QAIRT 2.49, 512x512, 400-row calibration captured by hooking unet.forward (so the 9-channel concat order and scaling are correct by construction), --act_bitwidth 16, per-channel weights, --bias_bitwidth 32.

Quantisation of the timestep path was checked against onnxruntime ground truth before release — 36/36 tensors within 0.997–1.34 of their true range, with /time_proj/Mul at 951.000 against a measured 950.984. That path is the one that silently destroys output when mis-ranged.

Both tiers then passed an on-device image gate (seed 12345, denoise 1.0, 20 steps):