File	Purpose
`unet.bin`	9-channel inpainting UNet, w8a16
`512x768.patch`, `768x512.patch`	zstd patches against `unet.bin` for the two non-square resolutions
`clip_v2.mnn`, `pos_emb.bin`, `token_emb.bin`	text encoder (MNN) + embedding tables
`vae_encoder.bin`, `vae_decoder.bin`	VAE, w8a16
`tokenizer.json`	CLIP tokenizer

`AbsoluteRealityInpaint_qnn2.49_min.zip` (1,002,848,600 bytes)

The same checkpoint compiled for small-VTCM hardware — dsp_arch v68, socModel 30, 2 MB VTCM — for chips that reject the _8gen2 context with Request feature vtcm size with value 8388608 unsupported.

sha256 a5b9db368c99e5edd6a4d8ebca095a219d92dbdb84e14fb639514e4ad54d4619

Same seven files minus the two resolution patches: those are zstd byte-diffs against the _8gen2 unet.bin and would reconstruct garbage here, so _min is 512x512 only. Its VAEs are v68 builds of their own, not the _8gen2 pair.

Both tiers are emitted from the same libmodel.so, so the quantisation is shared and only the graph config differs. Measured on an S25 Ultra (512x512, 20 steps, denoise 1.0, seed 12345), scored on the fraction of pure black/white pixels inside the mask:

tier	`dsp_arch`	VTCM	score	time
`_8gen2`	v73	8 MB	0.0144	3796 ms
`_min`	v68	2 MB	0.0092	9689 ms

_min renders indistinguishably from _8gen2; it costs about 2.6x the time and nothing else.

True 9-channel inpainting: conv_in takes [noisy latent 4 | mask 1 | masked-image latent 4], so the model sees the hole and its surroundings and can remove objects — unlike latent-blending mask support.

Compatibility

QAIRT 2.49.

_8gen2 is dsp_arch v73 (socModel 43) and wants 8 MB of VTCM, so it needs roughly Snapdragon 8 Gen 2 or newer — QNN contexts run forward only.
_min is dsp_arch v68 with 2 MB VTCM, which runs forward onto every HTP we know of and fits far more hardware.

⚠ VTCM is not implied by the marketing generation. The Snapdragon 8s Gen 3 (SM8635) is newer than an 8 Gen 2 and still lacks 8 MB — the "s" tier ships a cut-down HTP. Pick the archive by measured capability, not by chip age.

⚠ Both tiers are QAIRT 2.49 builds, so both carry contextBlobVersion 4.0.4 and an fp16 execution requirement stamped in by the compiler. A device without 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, never the fp16 one. Reported on SM8735.

Self-contained by measurement

Unlike the DreamShaper inpaint build, this ships its own CLIP and VAE rather than borrowing them. AbsoluteReality's text encoder is 0/196 tensors bit-identical against DreamShaper's and its VAE 0/248 (max abs diff 0.057). Sharing is a checkpoint-family property, and these are different families — borrowing would load without error and quietly degrade every generation.

Resolutions: 512×512, 512×768, 768×512 (_8gen2); 512×512 only (_min).

AbrahamPJ/absolutereality-inpainting-qnn

AbsoluteReality Inpainting — QNN / Snapdragon NPU

What's in the archive

AbsoluteRealityInpaint_qnn2.49_min.zip (1,002,848,600 bytes)

Compatibility

Self-contained by measurement

AbrahamPJ/
absolutereality-inpainting-qnn

`AbsoluteRealityInpaint_qnn2.49_min.zip` (1,002,848,600 bytes)