12-step Krea 2-distilled LoRA pushes Anima to photorealism, with full pipeline and fixes
Honest_Concert_6473 · reddit · 2026-11-03
The author retrained a Krea 2 Turbo-distilled LoRA on Anima with a photorealistic-caption dataset, making outputs consistently photoreal at CFG 1 / 12 steps. Trade-offs: crushed micro-details and inherited noise, fixable via a 0.5–0.7 denoise second pass into Krea 2 Turbo (same VAE) or standard img2img. Key finding: changing only the dataset completely flips the distilled model's style bias; training scripts are public.
More from Multimodal
- YUE2 issue: reference audio makes lyrics turn into mumbling instead of following the words — NefariousnessFun4043 · 2026-11-03
- Firecrawl × HeyGen MCP turns your website into brand videos with one prompt — toolstelegraph · 2026-11-03
- AI-generated 'GOOP sneakers' concept design goes viral with surreal slime aesthetic — Delahuntagram · 2026-11-03
- TalkPix ships remote MCP server with server-side spend limits for agents — No-Influence-161 · 2026-11-03
- AI Video Fails Double Feature: Sleepwalking Kid and the Runaway Coffee Machine — nastyalarm · 2026-11-03
- Runway launches Runway Dev, an API platform built for coding agents — runwayml · 2026-11-03