A Reddit user trained a Krea2 style LoRA locally on an RTX 5070 Ti 16 GB
y3kdhmbdb2ch2fc6vpm2 · reddit · 2026-07-24
First style LoRA trained for Krea2 on a local RTX 5070 Ti
A Reddit user says they trained their first-ever style LoRA for Krea2, after also making a character LoRA for Ciri from The Witcher 3.
- They used OneTrainer on an RTX 5070 Ti 16 GB with 32 GB RAM.
- Training setup included rank/alpha 16, 768 resolution, 0.0002 learning rate, batch size 1, accumulation steps 2, 60 epochs, AdamW, cosine schedule, and w8a8.
- The post links to a CivitAI model page and notes the LoRA can also produce more explicit outputs when combined with refusal LoRA setups.
More from coding & agent
- Codex Mac app fills in Samsung RMA forms and identifies the right drive — doodlestein · 2026-07-24
- Open-source Skill lets Codex compare Grok CLI and Kimi CLI in parallel — vista8 · 2026-07-24
- An agent can now pay for its own data using the x402 protocol — kleffew94 · 2026-07-24
- SkillHone stores an agent’s skill-edit history so later runs don’t relearn the same bugs — 机器之心 · 2026-07-24
- Codex configures a camera over USB, turning a bad UI into a scripted workflow — yacineMTB · 2026-07-24
- Open-source React library turns agent tool calls into reusable frontend UI — fuckysubreddits · 2026-07-24