Data Test: Z-Image Abliterated Encoder Does Not Uncensor
nathandreamfast · reddit · 2026-08-18
Addressing the debate over "abliterated" text encoders for Z-Image + Qwen3 4b, the author provides data-driven conclusions via benchmarking across 27 encoders using LPIPS and CLIP scores.
Key Findings:
- Semantic Integrity: Refused vocabulary shows near-identical cosine similarity (>0.9967) between heretic and base encoders. The encoder passes semantics intact and is not the censor.
- No Uncensoring: The base stack renders explicit content at a 100% rate; there is no render-stage censorship to remove.
- Visual Impact: Differences from encoder swaps (LPIPS < 0.19) are smaller than seed changes (0.52), indicating standard compression variance.
Related event: Testing 27 Encoders: Z-Image Abliterated Encoder Doesn't Lift Censorship(2 posts)→
More from Multimodal
- ImageNet should retire: DiT method rankings negatively correlate between ImageNet and text-to-image — YouJiacheng · 2026-10-16
- Open-source anatomy-reveal skill turns Claude into an exploded-view motion design studio — msharmas · 2026-10-16
- Flux 3 Klein vs Flux 2 Klein hands-on: the new base model 'barely usable' for T2I — Both-Rub5248 · 2026-10-16
- USTC paper uses speculative sampling to cut latency in interactive video generation — ustc · 2026-10-16
- 800,000 Gaussian splats run in WebGPU; Casado sees near-infinite scaling — generativist · 2026-10-16
- Nano Banana 2.1 image generation now live in Perplexity Computer for all users — ChrisUniverse · 2026-10-16