Skip to content

LoRAs and embeddings

Two kinds of small add-on file change what a model draws without replacing it.

LoRAs

A LoRA is a small adapter (usually tens to hundreds of MB) that teaches a model a style, a character or a concept. In Nightmare, LoRAs work on Anima (the Anima DiT models; a LoRA makes each render about a third slower), SD 1.5 Swap (an SD 1.5 checkpoint converted in npuforge as SD1.5 Swap) and SDXL Swap (an SDXL checkpoint converted in npuforge 1.0.10 or newer as SDXL Swap, with LoRA ticked; 1.0.11 offers LoRA as SDXL Swap's only feature).

FLUX.2, Z-Image, Qwen Image and Krea 2

These models take LoRAs too. A LoRA on them can make a render noticeably slower, and the LoRA list says so on these models. A LoRA made for FLUX.2 Klein 9B does not work on Klein 4B, and the other way round.

Download from CivitAI or Hugging Face

On an SD 1.5 Swap or SDXL Swap generator node, open LoRAs and tap Search online. The Get LoRAs sheet opens on the most popular LoRAs for that node's family.

Get LoRAs

  • SD 1.5 / SDXL at the top picks the family. SDXL includes Illustrious, Pony and NoobAI LoRAs, which all run on SDXL Swap.
  • Type to search. Results update as you type. You can also paste a CivitAI model link or a Hugging Face repo link to open that LoRA directly.
  • CivitAI / Hugging Face picks the site, and Popular ▾ sets the order (Popular, Top rated, Newest).
  • Tap a LoRA to see its versions (on Hugging Face, its files). Each one shows its size, base model and trigger words. Download saves it straight into your LoRAs. The bar shows progress and Cancel stops it. The file is checked against the hash CivitAI publishes, and the trigger words become the LoRA's note.

CivitAI needs your API key for most downloads. It's free: civitai.com → Account settings → API Keys. Paste it in the box shown in the sheet, or in Settings → Models → CivitAI. The key stays on your phone and is sent only to CivitAI. Show mature content (off by default) searches civitai.red, which includes NSFW LoRAs.

CivitAI search is blocked in some countries

In Australia, for example, CivitAI refuses searches, and the sheet tells you so. Pasting a CivitAI model link still works, and downloads still work with your key. You can also use Hugging Face or a VPN.

The pills at the top choose the family: SD 1.5, SDXL, Anima, FLUX.2 Klein, Z-Image, Qwen 2.1 and Krea 2 (scroll them sideways). Opened from a node, the sheet starts on that node's family.

Import

Settings → Models → LoRAs → Import, and pick the .safetensors file. It is listed with its size; the bin deletes it. The folder button in a node's LoRA list does the same.

Use

  1. Tap a FLUX.2, Z-Image, SD 1.5 Swap or SDXL Swap generator node.
  2. Under LoRAs, tick the ones you want. A checkbox per installed file.
  3. Set each one's strength with its slider (1.0 is the adapter's full effect), or tap the number beside it to type one, such as 0.35 or 1.25. On the DiT models (FLUX.2, Z-Image, Qwen Image, Krea 2, Anima) and Wan the slider runs from -2 to 2 and stops at 0 in the middle: below 0 pushes away from what the LoRA learned. On the SD 1.5 and SDXL Swap models it runs from 0 to 2, because their NPU cannot apply a negative weight.
  4. Run.

The LoRA picker on a generator node

You can stack as many as you like. A LoRA is applied when the render starts, so switching one on or off, or moving its slider, reloads nothing. It does cost render time, in proportion to its size: about 6% for a typical FLUX.2 adapter, around 40% for a large Z-Image one.

If the LoRA's description names a trigger word, put that word in your prompt.

In the prompt: <lora:name:0.8>

A prompt copied from CivitAI often names its LoRAs inline, as <lora:name:0.8>. Nightmare reads those tags: each is taken out of the prompt before it reaches the text encoder, and if the name matches a LoRA you have installed (with or without .safetensors), it is applied at that weight on top of the node's LoRA list. The tag wins if both name the same one. A tag for a LoRA you do not have, or any tag on a model that takes no LoRAs, is skipped and the run log says so. Tags do not count toward the prompt's token limit.

Type <lora: in a prompt and your installed LoRAs appear as chips under the box, filtered as you type: the ones already ticked on the node it feeds come first, then the ones made for its model. Tap one and it is written as <lora:name:1.0> with 1.0 selected, so you can type the strength straight over it.

Sort, filter and favourites

With two or more LoRAs, the list has a sort menu. Name, Newest, Most used (how often you tick it) or Size: and chips for each base model found among your files (SD 1.5, SDXL, FLUX, Z-Image, Qwen, Anima, Krea 2, Wan), the node's own first. The base is read from the file itself and shown under each name. If a LoRA is not recognised (or is filed under the wrong model), open its ⋮ and pick the right one under Model; Auto goes back to what the file says. Tap the star to make a LoRA a favourite: favourites come first. With eight or more, a search box filters by name. The LoRAs ticked on the node always stay at the top, in the order they are applied.

Folders

With dozens of LoRAs, put them in folders: open a LoRA's ⋮ and under Folder pick one of that model's folders, or New folder and type a name (Style, Anatomy, Effects). Folders show first in the list, closed, each with how many LoRAs it holds; tap one to open it. None takes a LoRA out again, and a folder with nothing in it disappears. The files stay where they are on the phone. While you search, every match shows, folders or not.

Wan's Motion transfer takes no LoRAs: it has no LoRA list, and <lora:…> tags in its prompt are skipped.

A LoRA trained for a slightly different version of the model may fit only in part. The picture still renders, and the run shows a warning naming the LoRA and how many of its parts were used (for example 840 of 1092): the rest of what it learned is missing from the picture.

Notes

The ⋮ beside each LoRA opens a note for it: its trigger words, the strength and settings it works best with, anything you would otherwise have to remember. Put the trigger words on the first line. A LoRA downloaded with Search online comes with its trigger words already filled in. In the note:

  • Copy copies the whole note.
  • Add to prompt adds the first line to the end of the prompt wired into that generator.
  • Replace prompt replaces that prompt's text with the first line.

A dot beside the ⋮ marks a LoRA that has a note. Notes are kept next to the LoRA files, so they stay when the app is updated.

Delete

Delete in a LoRA's ⋮ removes the file from the phone, after asking. If the LoRA was ticked on that node, it is unticked too. Any other node that uses it will refuse to run until you add it again or untick it. Settings → Models → LoRAs deletes LoRAs too.

On SD 1.5 Swap

An SD 1.5 Swap model has room for ONE rank-64 adapter per layer, so the LoRAs you tick are merged into it on the phone the first time that mix is used, about 10 s for one large LoRA or 25 s for two, and kept for next time. With a single LoRA the strength slider is free; changing the strengths of several makes a new merge. A LoRA trained above rank 64 is reduced to its best rank-64 version. Only the part of a LoRA that changes the image model applies; the part trained into the text encoder does not (true of every NPU conversion).

On SDXL Swap

The same, with SDXL LoRAs: an SDXL Swap model has 700 rank-64 slots, filled on the phone the first time a mix is used and kept for next time. LoRAs written with either layer naming (diffusers or SGM) are read. From npuforge 1.0.12 an SDXL Swap conversion can also keep ControlNet and IP-Adapter (nodes), as the default Illustrious XL Swap does; npuforge's SDXL Swap conversion takes about an hour on the phone.

Embeddings

An embedding (textual inversion) is a tiny file holding one trained concept. Often a style, or a "negative" such as a quality fix.

Settings → Prompts → Embeddings → Import, pick the .safetensors, and then name it in a prompt (or the negative prompt) to use it.