File	Type	Approx. size	Notes
`lens-f16.gguf`	`f16`	7.7 GB	Highest precision GGUF export
`lens-q8_0.gguf`	`q8_0`	4.1 GB	High quality quantization
`lens-q6_K.gguf`	`q6_K`	3.2 GB	Strong quality/size balance
`lens-q5_K.gguf`	`q5_K`	2.7 GB	Smaller balanced option
`lens-q4_K.gguf`	`q4_K`	2.2 GB	Smallest provided option

Usage

Example with stable-diffusion.cpp:

bash

./build/bin/sd-cli \
  --diffusion-model lens-q4_K.gguf \
  --llm /path/to/gpt-oss-20b-UD-Q8_K_XL.gguf \
  --vae /path/to/flux2_ae.safetensors \
  --vae-format flux2 \
  --cfg-scale 5.0 \
  --diffusion-fa \
  -W 1024 \
  -H 1024 \
  -p "A crystal dragon soaring through an aurora borealis sky, its entire body made of transparent faceted crystal refracting the green and purple aurora light into rainbow spectra, ice particles trailing from its wings, high fantasy digital art"

If you keep the same folder structure as stable-diffusion.cpp examples, a local command can look like:

bash

./build/bin/sd-cli \
  --diffusion-model models/Lens/lens-q4_K.gguf \
  --llm models/gpt-oss-20b-GGUF/gpt-oss-20b-UD-Q8_K_XL.gguf \
  --vae models/vae/diffusion_pytorch_model.safetensors \
  --vae-format flux2 \
  --cfg-scale 5.0 \
  --diffusion-fa \
  -W 1024 \
  -H 1024 \
  -p "A crystal dragon soaring through an aurora borealis sky"

The base Lens model is not the distilled Turbo variant, so the example uses the default stable-diffusion.cpp step count instead of the 4-step Turbo setting.

Download

Download one quantization:

hf download rootonchair/Lens-GGUF lens-q4_K.gguf --local-dir models/Lens-GGUF

Download all GGUF files:

hf download rootonchair/Lens-GGUF --include "*.gguf" --local-dir models/Lens-GGUF

Quantization Notes

The GGUF files were produced from the Lens diffusion transformer weights using stable-diffusion.cpp convert mode:

bash

./build/bin/sd-cli \
  -M convert \
  -m models/Lens/lens_bf16.safetensors \
  -o models/Lens/lens-q4_K.gguf \
  --type q4_K

The original sharded safetensors transformer was first merged into a single safetensors file before conversion.

No Q2 quantization is provided. In practice, Q2 is an extreme low-memory option and is more likely to noticeably degrade image quality. For most use cases, start with q4_K, q5_K, or q6_K; use q8_0 or f16 when quality is more important than size.

License and Attribution

This repository follows the upstream microsoft/Lens license. Please review the original model card and license before using these weights.

Original model:

microsoft/Lens

GGUF conversion:

rootonchair/Lens-GGUF