acojaco / dreamshaper-lcm-gguf

DreamShaper LCM (v7 and v8) - GGUF q5_1

Q5_1 GGUF quantizations of two LCM (Latent Consistency Model) DreamShaper checkpoints, for use with stable-diffusion.cpp . LCM models generate good images in very few steps (about 4 to 8) at low guidance, which makes them well suited to CPU and low-VRAM GPU inference.

These files are for stable-diffusion.cpp only . They do not work with llama.cpp, diffusers, or ComfyUI's native loaders.

Files

Both are SD1.5-architecture models and contain the UNet, text encoder, and VAE in a single file.

SHA256

Usage

sd-cli -m LCM_Dreamshaper_v7-q5_1.gguf \
  -p "portrait photo of a woman, detailed" \
  --sampling-method lcm --steps 4 --cfg-scale 1.5 --seed 42 -o out.png

sd-cli -m DreamShaper8_LCM-q5_1.gguf \
  -p "portrait photo of a woman, detailed" \
  --sampling-method lcm --steps 6 --cfg-scale 1.5 --seed 42 -o out.png

Recommended settings for both models:

• Sampler: lcm
• Steps: 4 to 8
• CFG scale: 1 to 2 (higher values tend to burn the image)
• Resolution: 512x512 is the native size for SD1.5; 512x768 or 768x512 also work

Older stable-diffusion.cpp builds name the binary sd instead of sd-cli .

How these files were made

Tested with stable-diffusion.cpp <version or commit> .

LCM_Dreamshaper_v7

1. Source: LCM_Dreamshaper_v7_4k.safetensors from SimianLuo/LCM_Dreamshaper_v7 (fp32).
2. Converted to fp16: <describe the method you used> .
3. Quantized:

   sd-cli -M convert -m <v7 fp16 file> -o LCM_Dreamshaper_v7-q5_1.gguf --type q5_1 -v
   

DreamShaper 8 LCM

1. Source: DreamShaper8_LCM.safetensors from Lykon/dreamshaper-8-lcm (already fp16).
2. Quantized directly:

   sd-cli -M convert -m DreamShaper8_LCM.safetensors -o DreamShaper8_LCM-q5_1.gguf --type q5_1 -v
   

Notes on quantization

• The converter quantizes mainly the large weight matrices. Convolution kernels, biases, norms, and other small tensors stay at their source precision (F16 here), so the file is larger than a pure 5-bit estimate.
• Going through fp16 first for v7 gives a much smaller file than converting from fp32 directly, with no meaningful quality cost for SD1.5 inference.
• Q5_1 is slightly more accurate than Q5_0 at about 9% larger size. Expect small differences from the original fp16 models, mostly in fine texture.
• The original LCM_Dreamshaper_v7 UNet has an extra guidance-conditioning tensor ( time_embedding.cond_proj ) that the v8 LCM does not. This is expected and not an error.

Sample images

Same prompt and seed, quantized output:

Credits

• DreamShaper by Lykon .
• LCM_Dreamshaper_v7 by Simian Luo , from the Latent Consistency Models project (Luo et al., 2023).
• Stable Diffusion 1.5 by RunwayML and CompVis / Stability AI, which these models derive from.

This repository only repackages and quantizes the above models. All credit for the original training belongs to their authors.

License

These files are derivatives of Stable Diffusion 1.5 based models and are distributed under the CreativeML OpenRAIL-M license. The license, including its use-based restrictions (Attachment A), applies to these files and to anyone you pass them on to. The full text is available at https://huggingface.co/spaces/CompVis/stable-diffusion-license .

You are responsible for using the models in line with that license and with applicable law. I am not the author of the original models and give no warranty of any kind.

If you redistribute these files, you must include the license and keep the use restrictions.