Bring your own model¶
Beyond the catalogue, you can use models you converted or downloaded yourself.
To import anything: tap + next to the Models title, optionally type a name, tap
Choose a file and pick it. The app reads the file and works out what it is: a converted
.zip, a checkpoint .safetensors or .gguf (and which model it is for), an upscaler .bin, or an
ESRGAN upscaler .pth or .safetensors (it lands under Upscalers).
It asks only when a checkpoint does not say which model it was made for. A LoRA or an embedding
picked by mistake is named, with where to add it instead.
SD 1.5, SDXL and Anima: a converted zip¶
These families run as models compiled for the NPU, so a normal .safetensors from CivitAI
cannot be used directly. It has to be converted first. Two ways:
- On the phone, with npuforge: it turns an SD 1.5
.safetensorscheckpoint into an NPU model with no PC involved. Its 9-channel inpainting exports run as real inpainting models, and its SD1.5 Swap exports appear under the SD 1.5 Swap family and take LoRAs and ControlNet per render (nodes). npuforge's SDXL Swap exports appear under the SDXL Swap family and take SDXL LoRAs per render (LoRAs); from npuforge 1.0.12 they can also keep ControlNet and IP-Adapter. A conversion has only the features ticked when it was made: the others are dimmed on its node, and tapping one says so. NPuForge 1.0.8 and later also marks v-prediction packages, which Nightmare enables automatically. - A zip made for LocalDream: converted models shared for LocalDream import unchanged.
Import the .zip with + (above). The app reads which family it is from the files inside.
If a package contains LocalDream's empty V_PRED marker, Nightmare launches the
native backend in v-prediction mode. Packages without it keep the existing epsilon
behavior.
FLUX.2, Z-Image, Krea 2, Anima and MOSS-Image: one .safetensors or .gguf¶
These build their model at load time from ordinary weights, so a fine-tune from CivitAI or
Hugging Face imports as-is. No conversion, no PC. Both .safetensors and .gguf work.
Import the file with + (above). The app reads the file's tensor names to tell Anima, FLUX.2 Klein (4B or 9B), FLUX.3 Klein, Z-Image, Qwen Image 2.1, Krea 2 and MOSS-Image apart, so an Anima file from CivitAI lands on the Anima page without you choosing anything.
On Family ▾ → Anima, both kinds of Anima are listed: first the ones on the DiT engine (any Anima file from CivitAI), then the ones converted for the NPU (zip).
- Klein 4B and Z-Image share the text encoder, VAE and tokenizer with the family's model, so an import costs only its own weights (plus up to 2.6 GB once, if you have neither family).
- Klein 9B and Krea 2 get their own copy of their text encoder (4.8 GB and 2.6 GB), copied from the built-in model if you have it, downloaded if not.
- Anima takes its text encoder and VAE (0.9 GB) from Anima Turbo 1.1 the same way. Anima 2B and 2.9B (40 blocks) both import; a 2.9B starts at 28 steps and CFG 4, as a base model wants; another size is refused. An
imported
.safetensorsruns as it is: about 20% slower than the built-in, which the app converts to Q8_0. A Q8_0.ggufruns at the built-in's speed. - MOSS-Image uses the text encoder, VAE and tokenizer of the MOSS-Image you have installed, so
an import costs only its own weights (without one, it downloads its own 5.6 GB). An import gets
the built-in's 8-step speed-up too (the same LoRA, read from the installed MOSS-Image, 8 steps
and CFG 1), shown as an 8 steps badge on its card. Switch it off with ⋮ → 8-step speed-up
for a fine-tune that wants its own steps: it then starts at 25 steps and CFG 3. A "Turbo" or
"8 step" file has the speed-up built in, so it never takes it twice (the app reads that from the
file's name) and starts at 8 steps and CFG 1. A 16-bit file that stops importing part way says
how much of it arrived: it is usually still downloading. A Q4_0
.ggufsuch as molbal's MOSS Image Turbo GGUF is the fastest MOSS-Image: about 40 s at 512x512 and 2.5 min at 1024x1024 on an 8 Elite (it is also in the catalogue as MOSS-Image Turbo). - FLUX.3 Klein uses the text encoder, VAE and tokenizer of the FLUX.3 Klein you have installed when it holds all three; otherwise it copies them from your other models or downloads them (2.7 GB). It starts at 16 steps and CFG 1, as the distilled Klein does.
- Qwen Image 2.1 gets its own copy of Qwen's text encoder, vision encoder and VAE (6.6 GB), copied from the built-in if you have it. An original Qwen-Image or Qwen-Image-Edit file is a different model and is refused.
Example: Moody Moss Mix (25 steps)
On its CivitAI page pick the fp8 file (V1.0 25 STEPS CFG3, about 4.8 GB); the int8 one imports too. Install MOSS-Image SFT first, then import the file with +, and switch ⋮ → 8-step speed-up off: it then starts at 25 steps and CFG 3, which is what it needs. About 3 min at 512x512 and 16 min at 1024x1024 on an 8 Elite.
Size and memory
A checkpoint bigger than about half your phone's RAM is marked likely too big for this
phone's RAM on its row: Android will probably close it while it renders. A smaller quant
(Q4_0, Q8_0 or fp8) of the same model is the fix. Q4_0, Q8_0, MXFP4 and fp8 are the formats
known to run on the NPU. K-quants (Q4_K_M, Q5_K, Q6_K...) run on the CPU, ten times
slower or more, and the run log says so when a render starts: pick the Q4_0 or Q8_0 upload.
int8, ConvRot and mxfp8 files
A .safetensors in ComfyUI's int8 format (int8, ConvRot) or in mxfp8 is converted to Q8_0 while it imports,
which the NPU runs. The result is about the same size and needs the same free space.
A Krea 2 fp8 .safetensors is converted the same way on phones before the 8 Elite, whose NPU cannot run fp8
directly: as fp8 its first block did not fit in the NPU's memory on an 8 Gen 3. A Krea 2 imported before 1.6.303
on such a phone needs importing again.
4-bit files
Files made for NVIDIA cards in ComfyUI's 4-bit formats (int4, nf4, nvfp4) and int8 .gguf files
are refused on import with a message: the engine cannot run them. Pick the .gguf (Q4_0 or Q8_0),
fp8 or int8 .safetensors version of the same model.
A folder copied onto the phone¶
If your models are in Download/Nightmare (Settings → Models folder),
you can copy a model's folder straight into Download/Nightmare/models/ with a file manager.
The app finds it the next time you open Models:
- A folder with LocalDream's marker file (
SDXL,ANIMA,KLEIN,ZIMAGE,QWEN_IMAGE_2_1,npucustomorfinished) is taken at its word. - A FLUX.2 or Z-Image folder with no marker is recognised from its weights.
A model's folder may also contain a config.json with its preferred prompt, steps, CFG and
scheduler, as LocalDream uses; the app reads it.