metadata
license: openrail++
base_model: stabilityai/stable-diffusion-xl-base-1.0
tags:
- stable-diffusion-xl
- text-to-image
- gguf
CyberRealistic XL v9.0 — GGUF
This repository contains the extracted components and GGUF quantizations of CyberRealistic XL v9.0 by Cyberdelia, intended for use with ComfyUI-GGUF.
The scripts and workflows used to generate these files are documented in insecure-erasure/safetensors-goodies.
Repository Structure
/
├── cyberrealisticXL_v90.gguf
├── cyberrealisticXL_v90-Q8_0.gguf
├── cyberrealisticXL_v90-Q6_K.gguf
├── cyberrealisticXL_v90-Q5_K_S.gguf
├── cyberrealisticXL_v90-Q4_K_M.gguf
├── cyberrealisticXL_v90-Q3_K_M.gguf
├── cyberrealisticXL_v90-Q2_K.gguf
└── safetensors/
├── cyberrealisticXL_v90_unet.safetensors
├── cyberrealisticXL_v90_clip_l.safetensors
├── cyberrealisticXL_v90_clip_g.safetensors
└── cyberrealisticXL_v90_vae.safetensors
SDXL Pipeline Comparison: Checkpoint vs GGUF + VAE Configurations
This section validates that different SDXL pipeline configurations produce equivalent results.
All three images were generated using the same prompt, seed, and sampler settings.
- Checkpoint: cyberrealisticXL v9.0 by Cyberdelia
- Prompt: ultra detailed bust portrait of a 28-year-old woman, medium shot, wearing turtle neck sweater, subject centered, natural framing, 50mm lens, realistic skin texture, detailed eyes, soft cinematic lighting, professional photography, high dynamic range, sharp focus, natural colors
- Negative prompt: lowres, blurry, oversharpened, jpeg artifacts, bad anatomy, extra fingers, deformed hands, cross-eye, plastic skin, overexposed, underexposed, watermark, text, logo
- Sampler: dpmpp_2m
- Scheduler: karras
- Steps: 30
- Denoise: 1.0
 |
 |
 |
|---|---|---|
| Full checkpoint — F16 Standard SDXL checkpoint with the VAE baked in at F16. Reference output. |
GGUF F16 + VAE extracted from checkpoint UNet in GGUF F16, VAE and CLIP extracted from the same checkpoint in F16. Pixel-perfect identical to the reference. |
GGUF F16 + madebyollin VAE fp16-fix — F16 Same as above but with madebyollin's VAE (originally FP32, converted to F16). |
The following table shows the difference between some of the K-quants available with llama-quantize.
 |
 |
 |
|---|---|---|
| Q8_0 | Q6_K | Q4_K_M |
Credits
This repository contains derived weights from CyberRealistic XL v9.0 by Cyberdelia, used under its original license.