Features

An overview of what Local Dream can do, with links to detailed pages.

Highlights

Three things most users miss on a first read — worth knowing about:

Custom Models — you are not limited to the built-in catalog. SD1.5 .safetensors can be imported directly in-app; pre-converted SD1.5/SDXL NPU .zip packages (community or your own host-side conversion) can be loaded the same way.
Inpaint — Fix Local Details on a Generated Image — repaint just one region of any generated image. The unmasked area is preserved exactly, and at save time the result is automatically composited back into the full uncropped original, with Laplacian Pyramid blending hiding the seam.
DMD2 Fast Models — community models with _dmd2 in the filename run roughly 4× faster with much less device heat (use LCM / CFG 1 / 8 steps). On a 16 GB+ device with Low RAM Mode off, SDXL generation drops to about 6–7 s per image. Try them first; if you need maximum quality, finish with an img2img refinement pass on the original model.

Prompting — txt2img, img2img, and inpaint modes; prompt weights (AUTOMATIC1111 syntax); token counter; textual inversion embeddings; inpaint canvas zoom and Laplacian Pyramid blending
Generation Parameters — schedulers (DPM++ 2M / SDE, Euler A / Euler, LCM) with optional Karras noise schedule, steps, CFG, denoise strength, seed, CLIP caching, and CFG = 1 optimization
Batch Generation — generate multiple variations in sequence with foreground-service support
Show Generation Process — preview intermediate diffusion results during sampling
Upscalers — built-in and custom 4× upscaler models (NPU)
SD1.5 NPU High Resolution — two-stage Highres.fix workflow with patch resolutions
SDXL Aspect Ratio — arbitrary aspect ratios on a single 1024×1024 model via centered-inpaint masking

Available Models — built-in catalog, pre-converted community models, and chip tier suffixes
DMD2 Models — _dmd2 community variants: ~4× faster, much less heat, with LCM / CFG 1 / 8 steps
Model Assets — custom checkpoint import, LoRA workflow, and model deletion
Tag Autocomplete — import CSV dictionaries for inline tag suggestions with translation search; imported embeddings are pinned to the top of the suggestion list
Seed Settings — reproducibility across CPU, GPU, and NPU modes

Download Sources — switch between Hugging Face, hf-mirror, or a custom URL; background downloads
Low RAM Mode — SDXL memory-saving behavior and speed tradeoffs
History — per-model generation history with filter/sort, batch save to gallery, copy parameters, 1-click img2img
Share Parameters — copy a generation's parameters to the clipboard and auto-import them on another device
App Settings — install variants, logging, and file manager
HTTP API Reference — drive Local Dream from scripts via the local HTTP backend

CLIP output caching — text-encoder outputs are persisted to disk and reused for any previously seen prompt, per model (cleared via Settings → File Manager; embeddings bypass the cache)
CFG = 1 optimization — on the NPU path, the unconditional UNet pass is skipped entirely, roughly halving step time
v_prediction models are automatically detected and handled correctly
Background downloads and batch generation continue via a foreground service when the app is in the background