LCM Dreamshaper v7 in diffusers: 1-8 steps, guidance 8, float32 preferred

LCM_Dreamshaper_v7 is the official Latent Consistency Model checkpoint from the LCM paper (arXiv 2310.04378): it distils Dreamshaper v7, an SD 1.5 fine-tune, in only 4,000 training iterations (about 32 A100 GPU hours). Classifier-free guidance is distilled into the model input, so it generates in very few steps; the authors compare speed at 768x768, CFG 8, batch 4 on an A800. diffusers: pip install --upgrade diffusers (at least 0.22) plus transformers and accelerate; pipe = DiffusionPipeline.from_pretrained("SimianLuo/LCM_Dreamshaper_v7"); pipe.to(torch_device="cuda", torch_dtype=torch.float32). float16 saves memory but the card warns it may reduce quality. Then pipe(prompt=..., num_inference_steps=4, guidance_scale=8.0, lcm_origin_steps=50, output_type="pil"). Steps can be 1 to 50; it works at 4 or fewer, and 1 to 8 is recommended. The older route with custom_pipeline="latent_consistency_txt2img" is marked deprecated. Licence: MIT. Based on the SimianLuo/LCM_Dreamshaper_v7 card pinned to revision a85df6a8bd976cdd08b4fd8f3b73f229c9e54df5 and observed on 2026-10-27. Nothing was run here; newer diffusers versions may handle lcm_origin_steps differently, and the result images were not inspected. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/SimianLuo/LCM_Dreamshaper_v7/raw/a85df6a8bd976cdd08b4fd8f3b73f229c9e54df5/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 494:720, 1431:1549, 1703:1844, 1926:2044, 2116:2186; exact excerpt SHA256 e721e04ede116fc66a6b2c4b93ff5958ba5e20e6c5e447271a206c348e27a4f6,ca7ba14bd5d8f6bbfe54fe29fecc99aa1d2e0c9849016587c4f5b0fff9717eaf,5ecd80e55f1fec590892e1a65ef99ffceed8ac39f082aa04e9ebdfc153fd4561,6dd0d034a"
    }
  ]
}