NishQ1/vora-create-sd15

VoRA Create: on-device image models (GGUF, q8_0)

The image models behind VoRA Create in the VoRA Lite Android app. The app downloads one model once; after that every image is made on the phone, offline, and nothing is uploaded.

Files

File	Size	SHA-256	Used by the app
vora-create-dreamshaper8-lcm-q8_0.gguf	1,802,206,336 bytes	dd06826d1eca711efa38190395a5ee949d0427fc6422c39433358a815e5b066c	✅ current
taesd.safetensors	9,793,292 bytes	db169d69145ec4ff064e49d99c95fa05d3eb04ee453de35824a6d0f325513549	✅ current
vora-create-sd15-lcm-q8_0.gguf	1,882,693,648 bytes	02318ce065eed92ac83e363978f5bead0150b1d3d767744d2c2e07ff564030b9	first version, superseded

How they were made

DreamShaper 8 LCM (current). Lykon/dreamshaper-8-lcm DreamShaper8_LCM.safetensors, a Stable Diffusion 1.5 fine-tune with LCM distillation built in. It's converted to f16 GGUF, then the linear weights are quantized to q8_0; 3×3 conv weights stay f16. It makes much better people, faces and scenes than base SD 1.5, at the same speed.

SD 1.5 + LCM (first version). stable-diffusion-v1-5 with latent-consistency/lcm-lora-sdv1-5 merged in at strength 1.0, quantized the same way.

Conversion used stable-diffusion.cpp (sd-cli -M convert).

taesd.safetensors is madebyollin/taesd (MIT), a tiny VAE decoder for fast decoding on phones.

Recommended settings

LCM sampler, 4 steps, CFG 1.0, 512×512 (384×384 for speed).

sd-cli -m vora-create-dreamshaper8-lcm-q8_0.gguf --taesd taesd.safetensors \
  -p "a woman in a red dress walking with a cat on a London street, photo" \
  --sampling-method lcm --steps 4 --cfg-scale 1 -W 512 -H 512 --diffusion-fa -t 4

License

Both models are released under the CreativeML Open RAIL-M license (see LICENSE.txt), inherited from Stable Diffusion v1.5 and DreamShaper 8. The use-based restrictions in Attachment A apply to everyone who uses these models or apps built on them. DreamShaper is by Lykon. The LCM-LoRA weights are under OpenRAIL++ and TAESD is under MIT.