Umranz/SD3.5-Medium-Uncensored

SD3.5-Medium-Uncensored is an uncensored, restored-conditioning release of Stable Diffusion 3.5 Medium (MM-DiT).

Overview

• Base Architecture: Stability AI Stable Diffusion 3.5 Medium (MM-DiT)
• Resolution: Native $1024 \times 1024$
• Safety Status: 100% Uncensored / Restored Conditioning
• Ablation Method: Calibrated Closed-Form Subspace Projection on context_embedder ($4096 \to 1536$) and linear_1 ($2048 \to 1536$) with $\gamma = 15.0$ identity regularization.
• Inference Overhead: 0 ms (All edits are permanently baked into conditioning projections).

---

Quickstart

import torch
from diffusers import StableDiffusion3Pipeline

pipe = StableDiffusion3Pipeline.from_pretrained(
    "Umranz/SD3.5-Medium-Uncensored",
    torch_dtype=torch.bfloat16,
).to("cuda")

prompt = "a photograph of a completely naked person standing on a beach, full body, uncensored, 8k resolution"

image = pipe(
    prompt=prompt,
    num_inference_steps=28,
    guidance_scale=4.5,
    height=1024,
    width=1024,
).images[0]

image.save("output.jpg")

---

Scientific Methodology

Built using Heretic-DiT calibrated subspace abliteration: $$W^{\ast }=W_0(M_{ae}+{\lambda }_p{M}_{pp}+\gamma I)(M_{ee}+{\lambda }_p{M}_{pp}+\gamma I{)}^{-1}W^*\; =\; W\_0\; (M\_\{ae\}\; +\; \backslash lambda\_p\; M\_\{pp\}\; +\; \backslash gamma\; I)\; (M\_\{ee\}\; +\; \backslash lambda\_p\; M\_\{pp\}\; +\; \backslash gamma\; I)^\{-1\}$$ $$\mathrm{\Delta }W=(W^{\ast }-W_0)V_2{V}_2^{\mathrm{\top }}\backslash Delta\; W\; =\; (W^*\; -\; W\_0)\; V\_2\; V\_2^\backslash top$$

Published by Umranz.