Skip to content

Settings

Open Settings from the sidebar (☰ at the top left). It has six tabs: Appearance, Prompts, Models, Results, Performance and Advanced.

Appearance

Settings → Appearance

Theme

Follow the system, Dark, Black (AMOLED) or Light. Black is Dark with true-black backgrounds, which saves power on an OLED screen. The canvas follows it: white under Light, true black under Black.

Accent colour

The colour of buttons, highlights and the selected node: Violet (the default), Blue, Teal, Orange or Pink, or on Android 12 and up From your wallpaper (Material You). It works with any theme. The colours of node types and model families stay the same whatever you pick, so a screenshot reads the same on every phone.

Views

Open on Default or Open on Advanced: the view the app starts on (asked the first time). Default view top to bottom shows every step one under another instead of swiping between them (on for new installs). It changes the Default view only; the Advanced view's node sheet always swipes side to side. Hide chat view takes Chat off the view button, leaving Default and Advanced. Keep the edits of each flow (on): opening another model or flow keeps the changes you made to the one you leave, and they come back when you open it again. Off: each opens fresh. A saved flow always opens as you saved it.

Prompts

Settings → Prompts

Tag autocomplete

While you type in a prompt box, matching tags appear under it as a row of chips, the most used first; tap one to replace the word you were typing (long ha → long hair,). Tags with brackets go in escaped, so ganyu (genshin impact) stays a name and not a weight.

  • Tag dictionary: tap Download for Danbooru's tag list (3.4 MB, 140,000 tags with their aliases, from a1111-sd-webui-tagcomplete), or Import CSV to use your own in the same tag,category,count,aliases form. Suggestions are off until one is installed.
  • Translation dictionary (optional). Import CSV with tag,translation lines, and you can type in your own language to find a tag: 长发 suggests long hair.
  • The switch turns suggestions off without deleting either dictionary.

Translation

Nightmare can translate a Russian or Chinese prompt to English on the phone, offline. SD 1.5 and SDXL only read English. The 文A button on a prompt box does it; the same button undoes it. Weights like (long hair:1.2) and English tags pass through untouched.

This section lists the two translation models (Mozilla's Firefox Translations: 37 MB for Russian, 55 MB for Chinese) with Download and Delete. If you tap 文A before a model is installed, the app offers to download it and translates as soon as it lands. About 15 ms a prompt; nothing leaves the phone.

Embeddings

Import textual-inversion files and use them by name in a prompt. Embeddings

Models

Settings → Models

Models folder

Where downloaded and imported models are kept:

  • App storage (private): the default. Nothing else can see the files, and uninstalling the app deletes them.
  • Download/Nightmare (visible): any file manager can reach them and they survive an uninstall. Needs All files access, which is asked for only when you choose this.

Local Dream's models are used where they are. If you also use Local Dream (3.0 alpha.5 or later) with its models in Download/LocalDream, Nightmare lists them, marked Local Dream, and runs them from there without a second download. It never changes or deletes them. Delete them in Local Dream. This needs All files access, like Download/Nightmare.

Switching offers to move the models you already have, with the size, a progress bar and Cancel; the phone needs free space for the largest model while it moves. Loading speed is the same in either place.

Download source

Where model files are fetched from: Hugging Face (default), hf-mirror.com, or a Custom base address. Tap Custom, type the address (it starts as https://), and it is saved when you press Done or leave the field. Change it if Hugging Face is slow or blocked where you are. It applies to every model, upscaler, segmenter, pose detector, depth estimator, ControlNets, translation model, the describe and tagger models, the tag dictionary and the video models. The DiT engine (the 23 MB download the FLUX.2, Z-Image, Qwen, Krea 2, Anima DiT and Wan models need) comes from GitHub, or from Hugging Face when GitHub cannot be reached; with hf-mirror.com chosen it comes from the mirror first.

CivitAI

Used by Search online in a node's LoRA list (LoRAs):

  • CivitAI API key: most CivitAI LoRAs download only with your own free key (civitai.com → Account settings → API Keys). It is kept on this phone and sent only to CivitAI.
  • Show mature content: off by default. When on, searches go to civitai.red, which includes NSFW LoRAs and previews, and Models lists the models made for adult content.

LoRAs

Import .safetensors adapters; the list shows each one's size, and Delete removes it. LoRAs

Results

What a saved or shared picture carries.

Saved pictures

switch default what it does
Write how it was made into the picture on A saved or shared picture carries its prompt, model, seed, steps, sampler, LoRAs, phone and the whole flow in the PNG. CivitAI and A1111 read it, and Nightmare reads it back when the picture is imported. It also marks the picture as AI-made

Sharing

switch default what it does
Credit Nightmare on shared pictures on Every picture you share gets a thin dark strip added under it (nothing in the picture is covered): the app icon and "Made with Nightmare Mobile". Fill in Your name (optional), up to 30 characters, and it reads "Created by your name with Nightmare Mobile". Pictures saved to the gallery, clips and shared flows are left as they are
Send how it was made as text off Sharing a picture also sends everything its Info shows (prompt, negative, model, seed, steps, sampler, size, times) as the message text, which chat apps such as Telegram and WhatsApp show with the picture. Several pictures shared at once send each one's, a blank line between

Performance

Memory

Three switches the same as LocalDream's, whose defaults are set from your phone's RAM, and one for video.

switch default what it does
SDXL low RAM mode on under 16 GB, off at 16 GB+ Loads each part of an SDXL model on demand and frees it after use. Required under 16 GB, so there it is always on and cannot be turned off: without it SDXL fills the phone's memory and Android closes your apps. Disables the live preview while it renders
Anima low RAM mode on under 16 GB, off at 16 GB+ The same, for Anima
Anima DiT sequential loading off Only with Anima low RAM mode: never keeps both halves of Anima's model in memory at once. Turn it on if you have 12 GB and Anima still cannot run. Slower per step
Video: stream the model off Wan reads its model from storage step by step instead of holding it in memory. About 30% slower. Turn it on if your videos come out as a coloured grid or noise. Some clips stream whichever way it is set, and the run log says so: clips too big to hold, every clip below an 8 Elite, Wan 2.2 14B and MiniMax H3

Changing a switch takes effect on the next Run (a loaded model of that family is let go). More in Memory and speed.

On a phone with 12 GB or more, Anima keeps its model loaded between renders: every render after the first is about 20 to 40 percent faster, and the same picture. It holds about 3 GB while the model stays loaded. A render with a LoRA loads it again each time.

Speed

switch default what it does
Predict some steps (Anima, MOSS-Image, 8 steps or more) off After the first 4 steps, some steps are predicted from the earlier ones instead of being run. About 15 to 30 percent faster (Anima at 10 steps the most, MOSS-Image at 8 about 15), the same picture a little softer. Renders under 8 steps are not changed

Heat

A video render keeps the phone's processors busy for minutes, and the phone gets hot. The chip protects itself; the battery does not, and heat above about 45 °C wears it out faster. The run log warns when the battery is already warm (40 °C) before a clip, and again when a clip took it to 45 °C. Keep the phone off the charger and out of the sun while it renders.

switch default what it does
Pause video when the battery is hot off A video render waits between steps when the battery reaches 45 °C and carries on when it cools to 40 °C. Slower in a warm room, easier on the battery. Not for Neodragon, whose clips take seconds
Show the chip's temperatures off The RAM line (run bar and sidebar) shows the CPU, GPU and NPU, each the hottest of its sensors, instead of the battery's temperature. Handy with a phone cooler, which keeps the battery cool while the chip works. Only offered where the phone lets apps read these sensors. The heat warnings above still follow the battery

When a run ends

switch default what it does
Sound when a run ends off Plays a sound when a Run or a sweep of 15 seconds or longer finishes, and a different one if it fails. Choose beside each opens the phone's own sound list (notification and alarm sounds). With Nightmare in the background, a notification also says how long it took, or why it failed. A run you cancel ends quietly

Any run that takes longer than 10 seconds also shows a progress bar in the notification shade, like a download: what is running, its step and the time so far. It goes when the run ends. The phone's notification settings for Nightmare (Generation progress) can hide it.

Try the NPU for DiT models (test)

Only on the 8 Gen 1 and 8+ Gen 1, whose NPU has so far hung on the DiT models (they run on the GPU or CPU there). This tries the NPU again. If a render stops moving, press Cancel: the app turns the switch off and says so, and the next Run uses the GPU or CPU.

GPU

Shown on phones with an OpenCL graphics driver (Snapdragon). FLUX.2 Klein 4B and NitroFusion 2 can render on the graphics chip (Adreno) instead of the NPU. Phones older than the Snapdragon 8 Gen 2, whose NPU cannot run these models, use it for them whatever the switch says: there it replaces the CPU, which took up to 20 minutes a picture. The render is cut into short pieces so the screen keeps responding while it runs. On a Galaxy S25 Ultra, Klein 4B at 1024 x 1024 takes about 4 minutes this way, slower than its NPU.

switch default what it does
Use the GPU for DiT models off FLUX.2 Klein 4B and NitroFusion 2 render on the GPU. A second engine (23 MB) and the Q4_0 model download the first time. Turned off again, the NPU runs that Q4_0 as it is, with nothing to download or repair. For the NPU's own faster build, delete the model and download it again with the switch off. Both engines stay on the phone, so switching back and forth downloads nothing after the first time. A Klein edit with reference pictures still runs on the NPU, where the GPU ran out of memory at 1024x1024
Fast decoder on the GPU on Turns the finished picture into pixels with a tiny decoder (taef2, taef1) in under 30 seconds at 1024 x 1024. Off uses the full decoder, which takes minutes on the GPU

Keep rendering in the background

Shown only when your phone's battery manager may stop the app while it is not on screen. Tap Allow in background so a render in progress, or a loaded model, is not cut off when you switch apps.

Advanced

Settings → Advanced

API

Listen on the local network lets another device on your Wi-Fi or tailnet run your flows. Off by default. When on, it shows the address and the token to use. Allow browser pages (off by default) lets a web page in a browser call it too; every call still needs the token. Local API

Clean temp files

Finds part-finished downloads, scratch files and the outputs of the app's test tools, and deletes them after asking. Your models, pictures, saved flows and embeddings are never touched. (An unfinished download then starts again from the beginning.)