A LoKr test on Krea2 looks a lot like LoRA, with decent early results
Beautiful_Egg6188 · reddit · 2026-07-28
LoKr on Krea2 behaves a lot like LoRA, with decent early results
The author trained a character adapter on 43 images and compared LoKr against LoRA on Krea2.
Setup details:
- captions were generated with Qwen3 VL 4B Instruct;
- each caption focused on composition, hair, expression, clothes, pose, and background;
- training used 10 repetitions × 43 images × 6 epochs, with saves every 430 steps;
- resolutions were 512 and 768;
- learning rate was 1e-4 with Automagic v2 optimizer;
- text embeddings and latents were cached.
On an RTX 4070 Super with 64 GB RAM, the author had to offload the text encoder fully and the transformer 75%. Their verdict: LoKr did not feel dramatically better than LoRA, but it produced decent results as early as 2k steps and behaved very similarly overall.
More from Multimodal
- Higgsfield opens up the full production blueprint behind its Originals — mhdfaran · 2026-07-28
- Designer turns an AI health-app concept into a playful redesign showdown — Tegadesigns · 2026-07-28
- OrcaDub 1.0 launches speech-to-speech video dubbing with 4.83 MOS — qinzytech · 2026-07-28
- Kinetic Abyss shows a 2D animation made with Stable Audio 3 and LTX 2.3 — Tadeo111 · 2026-07-28
- How to turn reference photos into a convincing impasto oil-paint workflow — Suspicious-Koala-570 · 2026-07-28
- A Midjourney style workflow uses GPT breakdowns to narrow hundreds of image references — beechinour · 2026-07-28