零存储危机隐私保护:您上传的证据影像仅在内存中进行即时神经分析,完成后立即彻底销毁。无日志、无云端备份。

Multimodal Diffusion Transformer威胁严重程度: 高度威胁

Stable Diffusion 3.5 & SDXL Detector: Latent Noise Forensics

Forensic analysis of Stability AI MMDiT architectures, SDXL latent representations, and LoRA weight artifacts.

使用 Sealed Rose 验证媒体真实性

实时神经网络检测 · 免费 · 无需注册 · 零存储 · 内存瞬时处理

验证深伪照片

Stable Diffusion (including SDXL and SD 3.5 Large) is the foundation of the open-source generative AI ecosystem. Open weights allow fraudsters to fine-tune custom checkpoints and LoRAs locally on consumer GPUs to generate targeted catfish personas, fake KYC credentials, and non-consensual imagery without guardrails. Sealed Rose inspects the unique latent space noise residuals and MMDiT attention maps left by Stable Diffusion models.

关键警讯与破绽特征

1Attention Map Leakage on Text and Badges

Logos, shirt typography, and background signage often display pseudo-letters or fractured glyph structures from unresolved text embeddings.

2Lobe and Ear Contour Warping

Ears frequently show fused cartilaginous ridges or uneven lobe attachment when rendered against complex hairstyles.

3LoRA Overfitting and High-Contrast Edge Halos

Locally fine-tuned LoRA models produce distinct contrast halos and repetitive grain patterns around high-frequency hair strands.

Stable Diffusion MMDiT Architecture Forensics

  • MMDiT Latent Noise Verification: Identifies the characteristic rectified flow and noise scheduler signatures of SD 3.5.
  • Frequency Domain Inspection: Uncovers high-frequency harmonic grid lines caused by standard VAE decoding bottlenecks.
  • LoRA & Checkpoint Signature Attribution: Measures style weight distribution to differentiate base models from custom community fine-tunes.

防范措施与应对步骤

步骤 1

Scan Image in Sealed Rose Forensic Suite

Upload the file to inspect VAE residual anomalies and detect Stable Diffusion generative checkpoints.

常见问题

Can Sealed Rose detect images created with custom SD checkpoints or Civitai models?

Yes. While custom checkpoints change aesthetics, they share the underlying SDXL or SD 3.5 VAE and MMDiT noise distribution detected by Sealed Rose.

支持 23 种语言: