SD 1.5 Swap models for the Snapdragon NPU (QNN)
The default SD 1.5 Swap models of Nightmare Mobile:
SD 1.5 checkpoints converted by npuforge (Convert as →
SD1.5 Swap) into a UNet whose LoRA (rank 64, 160 targets) and ControlNet residuals are graph
inputs, so LoRAs and ControlNets are chosen per render with no reconversion. The app downloads and
installs these from its Models tab; the ControlNets are in
AbrahamPJ/nightmare-sd15-controlnet-qnn.
| file | checkpoint | style | template |
|---|---|---|---|
absolutereality_v181_swap_v2_v79.zip |
AbsoluteReality v1.8.1 (Lykon) | realistic | Swap v2 — + IP-Adapter |
cuteyukimix_swap_v2_v79.zip |
CuteYukiMix Adorable (kemiaomiao) | anime | Swap v2 — + IP-Adapter |
anything_v5_swap_v2_v79.zip |
Anything V5 (Yuno779) | anime | Swap v2 — + IP-Adapter |
absolutereality_v181_swap_v79.zip |
AbsoluteReality v1.8.1 | realistic | Swap v1 (kept for older app builds) |
cuteyukimix_swap_v79.zip |
CuteYukiMix Adorable | anime | Swap v1 (kept for older app builds) |
Swap v2 (npuforge 1.0.7) adds IP-Adapter: each cross-attention layer's image-prompt K and V are
UNet inputs too, so a reference picture steers the render (IP-Adapter Plus / Plus Face, computed on
the phone's CPU from AbrahamPJ/nightmare-ip-adapter).
+3% per step when unused. It also widens the latent input's quantization window (v1: ±4.5, v2: ±8).
Chips: the _v79 builds were compiled on a Snapdragon 8 Elite, so they load on Hexagon v79
and newer (8 Elite, 8 Elite Gen 5). Builds for 8 Gen 2 / 8 Gen 3 (v73) and 888 / 8 Gen 1 (v68)
are planned. QAIRT 2.50 contexts; 512×512.
Contents (flat zip): unet.bin (the Swap UNet), vae_encoder.bin, vae_decoder.bin (QNN
contexts), clip_v2.mnn + embeddings + tokenizer.json (the text encoder, CPU), and
lora_targets.json — the order of the 160 LoRA inputs, which also marks the folder as Swap —
and, in v2, ip_targets.json (the 16 IP-Adapter layers).
Licence
Conversions of the named checkpoints, without retraining; each checkpoint's own licence and its author's permissions apply to its file (Stable Diffusion 1.5 derivatives: CreativeML OpenRAIL-M and its use-based restrictions). See the checkpoints' model pages for the authors' terms.