AbrahamPJ/nightmare-sdxl-controlnet-qnn

SDXL ControlNets as Qualcomm QNN context binaries (Nightmare Mobile)

Precompiled Hexagon NPU contexts of three SDXL ControlNets, for the SDXL Swap models in Nightmare Mobile. The app downloads the one that matches the phone's chip. They are not usable outside a QNN runtime.

file	source	chip
canny/controlnet_v75.bin	diffusers/controlnet-canny-sdxl-1.0	Snapdragon 8 Gen 3 (Hexagon v75)
canny/controlnet_v79.bin	same	Snapdragon 8 Elite (v79)
canny/controlnet_v81.bin	same	Snapdragon 8 Elite Gen 5 (v81)
depth/controlnet_v75.bin	diffusers/controlnet-depth-sdxl-1.0	v75
depth/controlnet_v79.bin	same	v79
depth/controlnet_v81.bin	same	v81
openpose/controlnet_v75.bin	thibaud/controlnet-openpose-sdxl-1.0	v75
openpose/controlnet_v79.bin	same	v79
openpose/controlnet_v81.bin	same	v81

Contract (one graph, model, all inputs 16-bit quantized): sample [1,4,128,128] NCHW, timestep [1], encoder_hidden_states [1,77,2048] (the first 77 tokens), text_embeds [1,1280], time_ids [1,6], controlnet_cond [1,3,1024,1024] NCHW in [0,1] → res_0 … res_9 NHWC at conditioning scale 1 (9 down-block residuals and the mid block).

How they were made: exported from diffusers with a plain attention processor, calibrated on 96 rows of real SDXL base 1.0 trajectories with the ControlNet active (photo and anime prompts, both CFG sides), converted with QAIRT 2.50 qnn-onnx-converter (w8a16, per-channel) and compiled per chip. Held-out rows on an S25 (v79) against fp32: canny 29.0 / 29.9 dB, worst residual cosine 0.9991; depth 29.8 / 30.0 dB, worst 0.9993; openpose 28.5 / 30.6 dB, worst 0.9989 (synthetic COCO-18 skeleton hints).

Licence: these are derivatives of the source ControlNets and carry their licences. Canny and depth: CreativeML Open RAIL++-M, including its use restrictions. Openpose: its source's card says "License: refers to the OpenPose's one" (CMU OpenPose, non-commercial) — non-commercial use only.