qwen-image-2512-sd

qwen-image-2512-sd — Qwen-Image-2512 + Lightning 8-step, curated for Enclave's sdcpp backend

Component-layout volume for enclave.host image generation: the Qwen-Image-2512 transformer with the lightx2v Lightning 8-step distillation LoRA merged into the weights, quantized to Q8_0 by the same stable-diffusion.cpp revision the fleet serves with. Generates in 8 steps at cfg 1.0.

Provenance

Every input pinned by revision; reproduction is tools/merge-lightning.sh in the image-generator app repo (streaming bf16 LoRA merge, then sd-cli -M convert --type q8_0):

• transformer: Qwen/Qwen-Image-2512 @25468b98e3276ca6700de15c6628e51b7de54a26 (bf16 shards)
• LoRA: lightx2v/Qwen-Image-2512-Lightning @a52649c9d0f6e1a248bff13f0df33bb8a2abdb52, Qwen-Image-2512-Lightning-8steps-V1.0-bf16.safetensors (720 rank-64 targets, alpha-scaled)
• text encoder: unsloth/Qwen2.5-VL-7B-Instruct-GGUF @68bb8bc4b7df5289c143aaec0ab477a7d4051aab, Qwen2.5-VL-7B-Instruct-Q8_0.gguf (unmodified)
• VAE: Comfy-Org/Qwen-Image_ComfyUI @46839d338df81ce625d5fae27d7e370314c0fbc9, split_files/vae/qwen_image_vae.safetensors (unmodified)
• quantizer: leejet/stable-diffusion.cpp @b5d812008eb7082a238fc589444544b3278187ae — the fleet's pinned serving revision

Smoke-tested at that same revision: 8 steps, cfg 1.0, 512×512, seed 7, euler — coherent output verified visually before upload.

Serving (sd.cpp)

sd-cli --diffusion-model diffusion.gguf --llm llm.gguf --vae vae.safetensors \
  -p "a red fox in snow, studio lighting" --cfg-scale 1.0 --steps 8 -W 1024 -H 1024

~34 GB peak VRAM at 1024px with tiled VAE decode. Component filenames are generic on purpose: Enclave's ENCLAVE_SD_*_FILE envs are node-global, so every SD volume on a node shares the same relative names.