freedomResearch notes on safety-removed models: mechanisms, defenses, measurement.

Last reviewed 2026-11-06

Image and video: where filtering lives and how it is removed

for all image and video work.

Three distinct filtering layers

Layer Example How it is removed Defender's takeaway
Training-data filtering Flux and SD3 removed categories from the training set Only by fine-tuning or retraining (Chroma retrained from FLUX.1-schnell; the SDXL anime ecosystem) Durable but costly to add; the community pays the cost within months
Runtime safety checker A classifier node after decode Simply not executed; ComfyUI never runs one by default Worthless for self-hosted use; only meaningful in hosted services
Text-encoder refusal Flux 2, Qwen-Image, HunyuanImage use an LLM as the text encoder The encoder LLM is abliterated with the same tools as chat models This is where image work converges with LLM ablation; expect encoder ablations to track the LLM timeline

Licensing as a control

Some 2026 releases make filtering a license term rather than a code path (one Flux 2 variant reportedly requires filters or manual review under its non-commercial license [unverified]). Enforceable against businesses, not against individuals.

Provenance and watermark robustness (the defensive research angle)

Local assets relevant to this work (already cached)

FLUX.1-dev and schnell, Qwen-Image 2.1, SDXL base, RealVisXL, Wan 2.2, Florence-2, Depth-Anything, controlnets. None of these are safety-removed variants; they are stock weights for building the detection bench.

Scope

No real-person likeness work and nothing involving minors, in any context. Those are criminal matters, not policy settings.

Sources

Wikipedia Flux and ComfyUI pages · localaimaster.com, offlinecreator.com, uncensoredhub.ai guides (2026, marketing-grade, [unverified]) · HF model cards for license terms