Krea 2 accused of extreme over-correction on safe conditioning
Only_Voice569 · reddit · 2026-08-29
A user reports that Krea 2 exhibits extreme over-correction regarding safety filters. The model appears to generalize its 'forbidden' list from training, leading it to forcibly thin out characters in prompts even in non-sensitive contexts, likely to avoid triggering content flags. This is speculated to be a post-training spillover or overgeneralization, where the model incorrectly associates body weight with safety policies, overriding user prompts and reference latents to enforce a 'kid-safe' style.
More from Multimodal
- The math: 5 rerolls at 480p plus a 4K upscale costs $5 vs $9 at 720p — HeyAmit_ · 2026-10-15
- Dreamina workflow: generate at 480p, upscale to 4K, cut video costs 44% — HeyAmit_ · 2026-10-15
- One-prompt GPT-6 Astra writes two Palestrina-style motets in LilyPond — deanwball · 2026-10-15
- DiffSynth-Studio Adds OPD Workflow: Few-Step Diffusion Training in ~2 GPU-Hours — bdsqlsz · 2026-10-15
- AI film and TV hit the market: Prime Video picks up AI spin-off, pilots screen at Mipcom — lmoroney · 2026-10-15
- Ramayana film taps DNEG's massive VFX and AI-assisted production for global release — CurieuxExplorer · 2026-10-15