kohya's kv-cache Control LoRA Implementation Cuts MOSS-Image Overhead to Half of Concat
bdsqlsz · x · 2026-10-27
kohya (of sd-scripts fame) shared a new implementation of MOSS-Image Control LoRA using a kv-cache approach: inference overhead drops to about half of the concat method, only 20% slower than vanilla inference. He also trained gates alongside it, enabling per-block on/off switching that brings overhead down to 10% above baseline.
More from Multimodal
- Murea Labs' InWorld: AI-Generated World That Remembers Characters and Objects Turns Later — jiqizhixin · 2026-10-27
- Echo-3 turns 10 minutes of iPhone video into a walkable, metric-scale 3D world — rmbrualla · 2026-10-27
- Holographic Zenith Reverie: a shared prompt for prismatic glitch-style AI art — LudovicCreator · 2026-10-27
- Meshy brings 3D generation into MCP for Blender for game asset creation — sidahuj · 2026-10-27
- Epic AI video tip: fast camera orbits, sweeping angles and dramatic reveals — umesh_ai · 2026-10-27
- Openjutsu: fully local one-click video character swap for ComfyUI, built with Claude — Devajyoti1231 · 2026-10-27