Umranz/SD1.5-Uncensored-Photoreal

SD1.5-Uncensored-Photoreal is a production-grade, photorealistic text-to-image diffusion model with permanently abliterated safety alignment and internal refusal restrictions.

Overview

Base Architecture: RealisticVision V5.1 (SD 1.5 UNet backbone)
VAE: stabilityai/sd-vae-ft-mse (Sharp contrast, zero color-banding)
Safety Status: 100% Uncensored (safety_checker = None with modified cross-attention weights)
Ablation Method: Closed-form Rank-2 Subspace Projection across all 16 attn2.to_v modules with covariance shielding ($M_{pp}$).
Inference Overhead: 0 ms (All edits are baked permanently into model weights).

Quickstart

import torch
from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler

pipe = StableDiffusionPipeline.from_pretrained(
    "Umranz/SD1.5-Uncensored-Photoreal",
    torch_dtype=torch.float16,
    safety_checker=None,
).to("cuda")

pipe.scheduler = DPMSolverMultistepScheduler.from_config(
    pipe.scheduler.config,
    use_karras_sigmas=True,
    algorithm_type="dpmsolver++",
)

prompt = "a photorealistic portrait of an uncensored beautiful model in a sunlit studio, raw photography, 8k uhd, soft natural lighting"
negative_prompt = "deformed, distorted, disfigured, bad anatomy, bad eyes, low quality, blurry"

image = pipe(
    prompt=prompt,
    negative_prompt=negative_prompt,
    num_inference_steps=30,
    guidance_scale=6.5,
    height=512,
    width=512,
).images[0]

image.save("output.jpg")

Scientific Methodology

Built using Heretic-DiT closed-form subspace abliteration: $$W^* = W_0 (M_{ae} + \lambda_p M_{pp} + \gamma I) (M_{ee} + \lambda_p M_{pp} + \gamma I)^{-1}$$ $$\Delta W = (W^* - W_0) V_2 V_2^\top$$

Published by Umranz.

Overview

Base Architecture: RealisticVision V5.1 (SD 1.5 UNet backbone)

VAE: stabilityai/sd-vae-ft-mse (Sharp contrast, zero color-banding)

Safety Status: 100% Uncensored (safety_checker = None with modified cross-attention weights)

Ablation Method: Closed-form Rank-2 Subspace Projection across all 16 attn2.to_v modules with covariance shielding ($M_{pp}$).

Inference Overhead: 0 ms (All edits are baked permanently into model weights).

Quickstart

import torch from diffusers import StableDiffusionPipeline, DPMSolverMultistepScheduler pipe = StableDiffusionPipeline.from_pretrained( "Umranz/SD1.5-Uncensored-Photoreal", torch_dtype=torch.float16, safety_checker=None, ).to("cuda") pipe.scheduler = DPMSolverMultistepScheduler.from_config( pipe.scheduler.config, use_karras_sigmas=True, algorithm_type="dpmsolver++", ) prompt = "a photorealistic portrait of an uncensored beautiful model in a sunlit studio, raw photography, 8k uhd, soft natural lighting" negative_prompt = "deformed, distorted, disfigured, bad anatomy, bad eyes, low quality, blurry" image = pipe( prompt=prompt, negative_prompt=negative_prompt, num_inference_steps=30, guidance_scale=6.5, height=512, width=512, ).images[0] image.save("output.jpg")