Dev releases standalone merge-extraction tool for Krea-2 models and LoRAs, no WebUI needed
aoleg77 · reddit · 2026-09-16
A Reddit developer released a standalone GUI tool for merging and extracting Krea-2 checkpoints and LoRAs, reviving much of SuperMerger's functionality without WebUI or Comfy.
Key features:
- Multiple merge methods: add difference, weighted sum, slerp, cosine A/B, TIES, DARE
- Simplified block merges targeting composition/character/style, with advanced controls
- Direct export to bf16/fp16/fp8/fp8 scaled/int8-convrot with no temporary files
- GPU or CPU merging, plus LoRA extraction from Krea-2 checkpoints
- Adaptive rank selection when merging multiple LoRAs, with a specified energy target, often yielding smaller LoRAs that retain capabilities
- Per-layer spectrum analysis and a merge advisor that heuristically suggests methods
The author also notes many community checkpoints weren't what they claimed — some were trained on Krea-2 Turbo then had Turbo weights deducted yet advertised as Raw. Packaged with Python + venv installers, DPI-aware UI with light/dark themes.
More from Multimodal
- Two friends made an 8-minute AI action comedy, playing both leads — Muted_Camp352 · 2026-09-16
- Stunning Three.js+TSL Coastal World 'Tideglass Cove' Released Open Source — techartist_ · 2026-09-16
- Grok Imagine rolls out beta text editing on generated images with font and color controls — testingcatalog · 2026-09-16
- ISMIR 2026 tutorials revealed: LLMs x Music and generative attribution featured — serrjoa · 2026-09-16
- Open-source music model YuE2 runs locally on just 4.5GB VRAM, one-click via Pinokio — cocktailpeanut · 2026-09-16
- Minimax H3 tip: blend latent noise to loosen overly literal reference images — neph1010 · 2026-09-16