Zero-Dataset Fine-tuning: Arthemy Suite Reshapes Krea-2 in Real-Time
ItalianArtProfessor · reddit · 2026-08-17
A Reddit user released Arthemy Krea-2 Tuner, an open-source ComfyUI suite that enables real-time model and CLIP editing by scaling transformer blocks and sub-tensors (e.g., Attention vs. MLP) without training or datasets.
The tool offers three levels of control:
- Tier 1 Block Tuner: Amplify or reduce whole block groups for general direction.
- Tier 2 Sub-Block Tuner: Adjust specific tensor types inside blocks for fine fixes.
- Tier 3 Chaos Tuner: Seeded random weight perturbations for exploration.
Calibrations can be saved as 1KB JSON presets. Currently calibrated for Krea-2 + Qwen3 but generalizable to other architectures.
More from Models
- Luna Max outperforms Sonnet in DeepSWE at 2.3% of the cost — reach_vb · 2026-08-17
- Qwen 3.8 27B Local Coding: 73 tok/s at 128K Context on RTX 4090 — julianharris · 2026-08-17
- DeepSeek launches V4 Pro priced up to 14x higher than V4 Flash — pstAsiatech · 2026-08-17
- Qwen 3.8 27b Local Coding Test: 73 tok/s at 128k Context, Usable Performance — julianharris · 2026-08-17
- Developer Praises Hermes Ecosystem: Everyone Building on Top — Saboo_Shubham_ · 2026-08-17
- Muse Glimmer Q8 Loops Badly, Fails at Coding in VSCode — Electronic_Back1502 · 2026-08-17