Available Models

Built-in Models

These models can be downloaded directly from inside the app.

Model	Type	CPU/GPU	NPU	Clip Skip	Source
SDXL Base 1.0	SDXL	❌	✅	-	HuggingFace
Illustrious v16	SDXL	❌	✅	-	CivitAI
AnythingV5	SD1.5	✅	✅	2	CivitAI
ChilloutMix	SD1.5	✅	✅	1	CivitAI
Absolute Reality	SD1.5	✅	✅	2	CivitAI
QteaMix	SD1.5	✅	✅	2	CivitAI
CuteYukiMix	SD1.5	✅	✅	2	CivitAI

TIP

Refer to each model's original page for recommended prompts, samplers, and CFG values.

SDXL is NPU-only

SDXL checkpoints are too large to run usably in CPU/GPU mode on a phone, so Local Dream only supports SDXL on the NPU path. SD1.5 supports all three paths (CPU / GPU / NPU).

If you are new to Stable Diffusion, see Generation Parameters for a beginner-friendly explanation of scheduler, steps, CFG, denoise strength, and seed.

Different models can prefer noticeably different settings. When possible, start from the example parameters on the model's official release page rather than assuming one sampler or CFG value fits everything.

Pre-converted Community Models

Community-maintained NPU model collections (for SD1.5 and SDXL):

• xororz/sd-qnn
• xororz/sdxl-qnn
• Mr-J-369
• YuuiKurata

The suffix on each file indicates the minimum chip tier required to run that model. A higher-tier chip can run any model at or below its tier; pick the highest suffix your device supports for the best performance.

SD1.5:

• _min — any Snapdragon NPU with Hexagon V68 or newer (non-flagship chips)
• _8gen1 — Snapdragon 8 Gen 1 / 8+ Gen 1
• _8gen2 — Snapdragon 8 Gen 2 or newer (2 / 3 / 8 Elite / 8s Gen 4 / 8 Elite Gen 5)

SDXL:

• _8gen3 — Snapdragon 8 Gen 3 or newer (3 / 8 Elite / 8 Elite Gen 5. This is the only SDXL suffix)

Some community models also include _dmd2 in the filename — these have DMD2 LoRA merged in and need their own parameter settings (LCM / CFG 1 / 8 steps). They are roughly 4× faster and run much cooler at a small quality cost. See DMD2 Models.

About Resolution Patches (SD1.5 NPU only)

SD1.5 NPU zips ship with a base 512×512 UNet plus optional resolution patches for extra resolutions (e.g. 768×512, 1024×768). A patch is a small .patch file inside the same zip (typically a few MB to ~10 MB, generated with zstd --patch-from against the base UNet) — not a separate download.

• Whether extra resolutions are available depends on the model: some packagers only ship 512×512, others ship the full set.
• The app applies the patch on the fly when you switch resolution from the model run screen. Switching reloads the UNet (1–2 s); CLIP and VAE stay loaded.
• Non-flagship chips (_min tier) do not support resolution patches.

Marker Files

Some .zip packages contain marker files at the package root that tell the app how to handle the model. They have no contents — only the presence of the file matters:

• SDXL — the package is an SDXL model. Without this, the app loads it as SD1.5.
• V_PRED — the UNet uses v-prediction instead of the standard ε-prediction. The app detects this automatically and switches sampler math accordingly.

Built-in models and pre-converted community packages already include the right markers; you only need to think about them if you are packaging your own.

Custom Models

You can also use your own checkpoints.

CPU/GPU mode (in-app)

Import any SD1.5 checkpoint directly in the app — conversion runs on-device. If you want to use a LoRA, merge it into the original checkpoint before import. Once the model has been converted and quantized, additional LoRA injection is not supported.

v_prediction models are automatically detected: if the model uses v-prediction instead of the standard ε-prediction, a V_PRED marker file must be present in the model directory. Built-in and correctly packaged conversions handle this automatically.

NPU mode (host-side conversion)

NPU models must be converted on a Linux/WSL host before they can be loaded. See:

• SD1.5 Conversion Guide
• SDXL Conversion Guide (experimental)

Technical Details

NPU

• SDK: Qualcomm QNN SDK on the Hexagon NPU
• Quantization: W8A16 static quantization
• Resolution: SD1.5 base at 512×512; additional resolutions (512×768, 768×512, 768×1024, 1024×768) available via patches for 8 Gen 1 and above. SDXL runs at a fixed 1024×1024 internally but supports arbitrary aspect ratios (long side = 1024, short side scaled) — see SDXL Aspect Ratio.
• Performance: Extremely fast inference

CPU/GPU

• Framework: MNN
• Quantization: W8 dynamic quantization
• Resolution: 128×128, 256×256, 384×384, 512×512
• Performance: Moderate speed, high compatibility