HyHorX/stable-diffusion-v1-5-lcm-hybrid

Stable Diffusion v1.5 – LCM Hybrid

Repository: HyHorX/stable-diffusion-v1-5-lcm-hybrid

---

Overview

This repository provides a hybrid Stable Diffusion v1.5 model with a focus on fast inference using LCM-style sampling.

Key characteristics:

• Base model: Stable Diffusion v1.5 (Diffusers format)
• UNet: LCM-compatible UNet (swapped)
• Scheduler: Latent Consistency Model (LCM) scheduler
• VAE: stabilityai/sd-vae-ft-mse
• Other components: Original SD 1.5 text encoder, tokenizer, feature extractor

This model is designed for low-step inference (2–8 steps) while maintaining SD 1.5 compatibility.

⚠️ Important
This is NOT a fully distilled LCM model.
It is a hybrid configuration intended for experimentation and fast generation.

---

What This Model Is

• ✅ Compatible with Stable Diffusion v1.5
• ✅ Uses an LCM UNet for fast sampling
• ✅ Uses LCM scheduler
• ✅ Improved stability with sd-vae-ft-mse
• ✅ Works directly with diffusers

---

What This Model Is NOT

• ❌ Not a full LCM distillation
• ❌ Not equivalent to official LCM checkpoints
• ❌ Not trained end-to-end with consistency loss

Only the UNet and scheduler are adapted for LCM-style inference.
All other components remain standard SD 1.5.

Recommended Settings

• Inference steps: 2–8
• Guidance scale (CFG): 1.0 – 2.0
• Scheduler: LCM only

Higher CFG or higher step counts may reduce output stability.

---

Example Usage (Diffusers)

from diffusers import StableDiffusionPipeline, LCMScheduler
import torch

pipe = StableDiffusionPipeline.from_pretrained(
    "HyHorX/stable-diffusion-v1-5-lcm-hybrid",
    torch_dtype=torch.float16,
)
pipe.scheduler = LCMScheduler.from_config(pipe.scheduler.config)
pipe = pipe.to("cuda")

image = pipe(
    prompt="a cinematic cyberpunk city at night",
    num_inference_steps=4,
    guidance_scale=1.5
).images[0]

image.save("result.png")

Notes on Quality

Optimized for speed, not maximum fidelity

Best results with LCM-style low-step sampling

Quality may differ from standard SD 1.5 at high steps

The use of sd-vae-ft-mse improves latent decoding stability compared to the original SD 1.5 VAE.

License & Credits

Base model: Stable Diffusion v1.5

VAE: stabilityai/sd-vae-ft-mse

UNet: LCM-compatible variant

Please follow the original Stable Diffusion license and Hugging Face usage policies.

Disclaimer

This repository is provided as-is for research and experimentation. It is clearly labeled as a hybrid model to avoid confusion with fully distilled LCM models.