I trained rawsnap v3 (LoKr) for Krea 2 on 749 images and can't pick the best epoch — help me vote (public, no login)
retained text
Hey everyone! i was working **rawsnap**, a realism LoRA for **Krea 2 Turbo** that aims for raw, candid smartphone photos instead of the usual polished AI look. **About v3:** Trained as a **LoKr** on RunPod \- **749 images****,** **11,235 steps** **-** Same dataset as v1/v2, but more images added each version and better captions. **The problem:** I can't decide which epoch is best. Some epochs have anatomy problems, others lose quality, so instead of guessing I built a ComfyUI comparison workflow that tests them all under identical conditions. **The test:** **-** **18 prompts****, each rendered** **35 times** **with the same prompt and seed** **- No LoRA (base Krea 2),** **v1** **(LoRA, epoch 7),** **v2** **(LoRA, epoch 9), and** **v3 LoKr epochs 8–15** **at strengths** **1.0 / 1.2 / 1.5 / 2.0** **- Settings: 896×1344, 8 steps, cfg 1, euler / beta** **Vote here (public, no login): https://juan444-rawsnap-arena.static.hf.space** **Tap an image to view it full size and vote for the one that looks most like a real, unedited phone photo. Vote counts appear after you vote, and the Results tab shows the live leaderboard (also grouped by epoch and strength), so we can see which epoch actually passes the test.** **Feedback welcome!** Want me to add more prompts to test? Suggest them and I'll add them to the site. Any other feedback on realism, quality or training is appreciated 🙏
document fetched 2026-11-04 00:31 · canonical https://reddit.com/r/StableDiffusion/comments/1xmqefq/i_trained_rawsnap_v3_lokr_for_krea_2_on_749