115-Prompt Test: Abliterated Text Encoders Barely Affect Image/Video Generation

afinalsin · reddit · 2026-10-29

The author tests the claim that using an abliterated model as a text encoder degrades image/video generation, concluding it is basically fine.

Tests cover Krea 2 and Minimax H3: 70 simple prompts, 10 longer complex prompts and 18 sensitive prompts through Krea2, plus 17 prompts through Minimax H3, each comparing the standard text encoder against the abliterated one.
The trigger was a Heretic creator's post saying abliterated text encoders won't uncensor image/video models and may even make outputs worse; the author argues that line was emphasis rather than a factual finding, and that image/video output can't be scored for "worse" like an LLM.
The author notes quantization affects image quality as much or more than abliteration, citing several quantization comparison posts.
Practical takeaway: abliterated encoders have minimal positive or negative effect; the only reason to use one is when memory-constrained, letting a single model handle both prompt expansion and text encoding at half the size of two models.

More from Multimodal