Krea 2 Turbo 4-step distill LoRA retrained to 133K steps with cleaner detail, less ghosting
TimeTruth2490 · reddit · 2026-11-05
lvladikov released a new 133K-step checkpoint of the Krea 2 Turbo 4-step distillation LoRA, replacing the September 78K 'final' release after 55K more training steps with five critic heads (up from one). Against the 8-step teacher across 15 test prompts at 12 sizes: grid artifacts dropped from 1.09–1.11× to 0.96–0.98×, ghosting/doubled edges improved at 10 of 12 sizes, flat-area grain fell from 1.08× to 0.99×, and over-sharpening at 1440px went from 1.18–1.20× to 1.01–1.04×. Training cost about one day on an H200 (RunPod) plus two days on an RTX 3090. The author notes the 2/3/4-step LoRAs also boost per-step quality on third-party Krea checkpoints that ask for more steps (e.g. the 2-step LoRA at 6–8 steps rivals 12–20 steps without it), and the 3-step LoRA was just updated to 66K.
More from Multimodal
- Tencent Hunyuan Hy4 builds interactive 3D objects with working parts from a single prompt — mhdfaran · 2026-11-05
- Quixel makes Megascans free again and allows AI training on its 3D scan library — VraserX · 2026-11-05
- One prompt builds an interactive 3D alpine valley with Tencent Hunyuan Hy4 — no models, no textures, all code — Aiden_Tech_Ai · 2026-11-05
- A total beginner made an ad in under 10 minutes by pasting one link into Claude — PrajwalTomar_ · 2026-11-05
- 10 open-source GitHub repos for voice AI: cloning, TTS, and phone-call agents — victor_explore · 2026-11-05
- Generative world demo shows riverside scene with motorbike, bow and fishing rod — wavefnx · 2026-11-05