sd-3.5-small

⚠️ Disclaimer: This is not an official model. It is intended for research and experimental purposes only. This is also cloned from https://huggingface.co/kpsss34/Stable-Diffusion-3.5-Small-Preview1/tree/main

NOW ComfyUI Custom nodes is created!!

• https://github.com/kpsss34/ComfyUI-kpsss34

Stability AI Community License, not MIT/Apache/GPL. It's a proprietary license with revenue-based restrictions.

Say Hi!

Dear All,

I’ve built on top of the SD3.5 model to improve both performance and efficiency. The original base model included several parts that used more resources than necessary. Some of the bias issues also came from DIT, the main image generation backbone.

I’ve made a few key changes — most notably, cutting down the size of TE3 (T5-XXL) by over 99%. It was using way too much power for what it did. I still kept the core features that matter, and while the prompt interpretation might be a little less powerful, it’s not by much, thanks to model projection and distillation tricks.

Personally, I think this version gives great skin tones. But keep in mind it was trained on a small starter dataset with relatively few steps, just enough to find a decent balance.

If you'd like to support future versions or want me to make something similar, feel free to suggest or support me here: 👉 https://buymeacoffee.com/kpsss34

Thanks, and enjoy using it!

Comparison Models

Prompt

• a woman with long brown hair and a fair complexion is captured in a close-up, eye-level shot, seated on a dirt path amidst a cluster of vibrant red flowers. she's dressed in a rust-colored spaghetti-strap dress, adorned with a flower pinned to her left cheek. her left hand rests on her right thigh, while her right hand rests on her left knee. her gaze is directed towards the camera, her lips slightly parted, her long, dark hair cascades down her back.
• eye-level shot, a dark-haired man, clad in a navy blue suit, stands on a set of blue stairs. the stairs, constructed from blue bricks, are interspersed with white bricks, adding a touch of contrast to the scene. the man's attire is a three-piece suit, consisting of a jacket, a vest, and a pair of pants. the jacket is buttoned up to the neck, while the vest is unbuttoned, revealing a white shirt underneath. the man's hair is neatly styled, with a goatee and mustache adding character to his appearance.
• a serene outdoor shot on a cloudy day, featuring a body of water and a forest in the distance. the water is a deep, dark blue, reflecting the sky and clouds in the distance. the foreground is dominated by a forest of tall, slender trees, their green leaves a vibrant contrast to the dark blue water. the sky is a canvas of dark, ominous clouds, their dark gray hues adding a sense of mystery to the scene. the sun is barely visible in the top right corner of the frame, casting a warm glow on the trees and the water.
• captured at eye-level on a radiant sunny day, the photograph showcases the picturesque landscape of jokulsarlon, a glacial lagoon in iceland. the foreground is dominated by a rugged shoreline, punctuated by large, jagged rocks. the water, a vibrant blue, is chock-full of icebergs, some of which are floating on the surface, while others are submerged. the background is a stark contrast, with snow-capped mountains towering over the scene. the sky, a deep blue canvas, is dotted with wispy white clouds, adding depth and contrast to the scene.

Resolution : 768*768

Step : 40

How to use

1.dowload pipeline my github : https://github.com/kpsss34/ComfyUI-kpsss34.git

• pipeline : pipeline_stable_diffusion_3_S.py

• inference_file : inference.py

2.Keep pipeline files and inference files in the same folder.

3.python inference.py

import torch
import time
from pipeline_stable_diffusion_3_S import StableDiffusion3SPipeline

seed = 43
generator = torch.manual_seed(seed)

pipe = StableDiffusion3SPipeline.from_pretrained(
    "kpsss34/Stable-Diffusion-3.5-Small-Preview1",
    torch_dtype=torch.bfloat16,
    use_safetensors=True,
).to("cuda")

filename = f"./img_{seed}_{int(time.time())}.png"

image = pipe(
    prompt="",
    negative_prompt="bad hands, bad finger, worst quality, low quality, jpeg artifacts, cartoon, painting, doll, ugly, disfigured, deformed, mutated, extra limbs, extra fingers, missing fingers, long neck, bad anatomy, bad proportions, unrealistic face, cloned face, blurred, watermark, text",
    num_inference_steps=40,
    guidance_scale=5.0,
    width=768,
    height=768,
    generator=generator,
).images[0]

image.save(filename)
print(f"Saved as {filename} (seed: {seed})")