Skip to content

Models

A model (or checkpoint) is the trained network that actually draws. Nightmare ships without any. You download the ones you want from Models, and each is downloaded once and kept.

Models come in families. A family decides how prompts are written, what sizes it draws and which phones can run it; inside a family, models differ in style.

The families

family models size each draws style needs
SD 1.5 6 ~1 GB 512–1024 Fast; realistic, anime and general mixes Snapdragon 888+
SD 1.5 Swap (experimental) AbsoluteReality, CuteYukiMix, Anything V5, or your own npuforge conversion ~1.3 GB 512 LoRAs, ControlNet (canny, depth, openpose) and an IP-Adapter reference picture chosen per render defaults: 8 Elite+; your own conversion: the phone that made it
SDXL 10, and 2 in Community (8 steps, DMD2 merged) ~3.5 GB 1024 Sharper, better composition 8 Gen 3+
SDXL Swap Illustrious XL, Juggernaut XL, 2 in Community, or your own npuforge conversion ~3.9 GB 1024 LoRAs, ControlNet (canny, depth, openpose) and an IP-Adapter reference picture chosen per render defaults: 8 Elite+; Community: 8 Gen 3+; your own conversion: the phone that made it
Anima 9 ~4.3 GB 1024 Anime; slow (about 80 s a picture) 8 Gen 3+
FLUX.2 Klein 4B, Klein 9B 6.7 GB, 10.7 GB 512–2048 Understands sentences; edits photos. 9B: more detail, slower, starts at 768 8 Gen 3+ (on 8 Gen 3 the 4B is a Q8_0 GGUF: 6.9 GB in all); the 9B also on 8 Gen 2
Z-Image Turbo 8.8 GB 512–2048 Photorealistic, fast for its size 8 Gen 2+ (on 8 Gen 2 and 8 Gen 3 it is a Q4_0 GGUF: 6.3 GB in all)
NitroFusion 2 1 step (2 if you want) about 9.8 GB (7.2 GB beside Z-Image Turbo: it shares its encoder and decoder) 512–2048 Built on Z-Image Turbo: a photorealistic picture in one step, about 20 s at 1024 on an 8 Elite. Tried with text to image so far 8 Gen 2+ (below the 8 Elite, and on the GPU, it is a Q4_0 GGUF: 3.5 GB, 6.1 GB in all; on an S25 Ultra's GPU 2 to 2.5 minutes at 1024)
Qwen Image 2.1, 2.1 Turbo 10.8 GB (16 GB phone: 13.8 GB FP8); Turbo 10.8 GB 512–2048 Follows long instructions, legible text; edits photos; slow. Turbo: 8 steps, about 2.5 times faster 8 Gen 3+ (8 Gen 3 always takes the 10.8 GB build)
Anima DiT Turbo 1.1 3.2 GB (4.2 GB download, converted on the phone) 512–2048 Anime; about a minute at 1024 8 Gen 1+ (on the 8 Gen 1 it runs on the CPU, about 15 to 20 minutes at 512)
Anima DiT 2.9B (preview) 4.2 GB (6.8 GB download, converted on the phone) 512–2048 Anime, a larger model with a built-in turbo LoRA: 8 steps at CFG 1, about 95 s at 1024 on an 8 Elite 8 Gen 1+
Krea 2 Turbo 9.5 GB 512–2048 Text to image only. On a 12 GB phone it renders in parts: about 3.5 to 4 min at 1024 12 GB RAM, and 8 Gen 3+
MOSS-Image SFT, Base and Turbo, 8 steps about 10 GB (Base 4.4 GB, Turbo 3.3 GB more beside SFT) 512–2048 Text to image only. Careful and detailed, with an 8-step speed-up applied every time: about 35 s at 512 and 2 min at 1024 on an 8 Elite. Fewer than 8 steps leaves blocks in dark areas (the run warns): for 4, use the 4-step model 8 Gen 3+
MOSS-Image SFT, 4 steps about 10 GB (1.2 GB more beside SFT at 8 steps) 512–2048 Text to image only. The same model with a 4-step speed-up: about 20 s at 512 and 1 min at 1024 on an 8 Elite 8 Gen 3+
Ideogram 4 8 steps about 11.6 GB 512–2048, starts at 768 Text to image only. The best at readable text in a picture (signs, posters, labels), with an 8-step speed-up applied every time; about 5.5 min at 768. Non-commercial use only 12 GB RAM, and 8 Elite+
FLUX.3 Image Klein, 16 steps about 6.2 GB (3.9 GB with Krea 2 installed: it shares Krea 2's text encoder) 512–2048 Text to image and Image edit. Black Forest Labs' newest model: sharp detail, good prompt following, readable text; about 45 s at 512 and 2 min at 1024. FLUX Kommunity License 8 Gen 3+
FLUX.1 dev, 20 steps about 10.8 GB 512–2048 Text to image only. Black Forest Labs' FLUX.1 dev: about 4 min at 512 and 6 min at 1024 on an 8 Elite. Non-commercial use only 8 Elite+, 12 GB
Fibo 2.0 Edit Turbo, 4 steps about 8.4 GB (6 GB with Krea 2 installed: it shares Krea 2's text encoder) 512–2048 Wire a picture into the Image edit flow and say what to change: Bria's 8B editor keeps the rest of the picture and writes text on signs. With no picture, text to image too (about a minute at 896x1600 on an 8 Elite). Non-commercial 8 Gen 3+, 12 GB
Iris 3B Turbo 5.8 GB (6 GB download, converted on the phone) only its trained sizes, about 1 MP: 1024x1024, 896x1200, 896x1344, 752x1344 and the turned versions Text to image only. Draws pixels directly, with no VAE; about 2.5 min at 1024 8 Gen 2+

Qwen Image 2.1 and its Turbo each keep their own copy of the text encoder and decoder (6.6 GB), so installing both takes about 21.6 GB.

Anima and Anima DiT share one page in Family ▾ → Anima: the DiT ones first, then the ones converted for the NPU.

The first model you download also fetches the NPU runtime for your phone's chip (about 28 MB, once); it stays through app updates. On an 8 Elite Gen 6 the DiT models run with a Gen 6 build of the engine's NPU code.

Besides picture models, Models also holds:

  • Upscalers: small models that enlarge a picture. Upscale
  • Video: the text- and image-to-video models, one 8.6 GB download, Wan 2.2 (6.25 GB), slower and better, and MiniMax H3 (17.5 GB), clips with sound, and MiniMax H3 Reference (14.4 GB more), a clip that keeps a picture's subject. Video

The Models tab, SD 1.5 family

Community models (a tab on the SDXL and SDXL Swap pages, and a badge on their rows) were converted by people in the community with npuforge on an 8 Gen 3, so they run on the 8 Gen 3 and newer, where our own SDXL Swap conversions need an 8 Elite. They are downloaded unchanged from their converters' pages.

Which to start with

  • Any phone: AbsoluteReality (realistic) or AnythingV5 (anime), SD 1.5. Small and quick.
  • To inpaint: add AbsoluteReality Inpaint, a dedicated inpainting model.
  • 8 Gen 3 and up: an SDXL model for sharper, larger pictures.
  • 8 Gen 3 and up: FLUX.2 Klein. Prompts in plain sentences, and photo editing.

In this section