Realism LoKr for Krea 2 trained on 749 images, blind voting arena picks the best epoch
Wide_Director_8897 · reddit · 2026-11-03
The author released rawsnap v3, a realism LoKr for Krea 2 Turbo targeting raw, candid smartphone-photo looks. Trained on RunPod with 749 images over 11,235 steps, it builds on the v1/v2 dataset with more images and better captions. To pick the best epoch, they built a ComfyUI comparison workflow: 18 prompts, each rendered 35 times with identical seeds, covering base Krea 2, v1 (epoch 7), v2 (epoch 9), and v3 epochs 8–15 at strengths 1.0/1.2/1.5/2.0 (896×1344, 8 steps, cfg 1, euler/beta). A public no-login voting arena shows a live leaderboard grouped by epoch and strength.
More from Multimodal
- RL for Diffusion Models: A Deep Dive on Why Rewards Beat DPO Pairs — kastnerkyle · 2026-11-03
- GPT image 2.0 demo: one prompt turns coffee photos into isometric shop illustrations — op7418 · 2026-11-03
- A 17-second AI car ad built story-first: script, then storyboard, then generation — EdenMakers-ai · 2026-11-03
- CapCut Multi-Track Workflow for AI Fashion Try-On Videos: Sync Outfit Changes to the Washing Drum — Aiden_Tech_Ai · 2026-11-03
- ComfyUI Workflow Generates 1080p Video with MiniMax H3 on Only 6GB VRAM — cgpixel23 · 2026-11-03
- Midjourney Recipe: 5 Weighted --sref Styles for Cinematic Light and Motion — michaelrabone · 2026-11-03