Nightmare — SDXL Swap models
Two SDXL checkpoints converted for the Qualcomm Hexagon NPU (HTP v79, Snapdragon 8 Elite and newer) by npuforge as SDXL Swap v2: the adapters below are graph inputs, chosen per render without reconverting. Any SDXL LoRA (rank ≤ 64); 462-token prompts (six CLIP chunks), 1024×1024. Downloaded and used by the Nightmare app.
| file | source checkpoint | per render | licence |
|---|---|---|---|
illustrious_xl_v1_swap_v3_v79.zip |
Illustrious XL v1.0 — OnomaAI Research | LoRA, ControlNet (canny / depth / openpose), IP-Adapter (Plus / Face) | CreativeML Open RAIL++-M |
juggernaut_xl_ragnarok_swap_v2_v79.zip |
Juggernaut XL Ragnarok — RunDiffusion | LoRA | CreativeML Open RAIL-M, with RunDiffusion's terms (no paid API deployment without their licence) |
illustrious_xl_v1_swap_v2_v79.zip |
Illustrious XL v1.0 — superseded by v3, kept for app versions before 1.6.122 | LoRA | CreativeML Open RAIL++-M |
Each zip holds the QNN context binaries (UNet, VAE encoder/decoder), the MNN CLIP-L / CLIP-G
encoders, the tokenizer and embeddings, lora_targets.json (and ip_targets.json where
IP-Adapter is kept), swap_features.json (what the conversion takes) and the 462_masked_v1
context marker. The ControlNets and IP-Adapter heads are separate downloads:
nightmare-sdxl-controlnet-qnn,
nightmare-ip-adapter. The use
restrictions of each source licence apply to these converted weights and to anything made with
them. Converted on a Samsung S25 Ultra: v2 on 2026-10-04/05 (npuforge 1.0.11), Illustrious v3
on 2026-10-07 (npuforge 1.0.12).