Most LoRAs only store low-rank adapters and therefore cannot reproduce normalization or modulation changes made by a full fine-tune. Kroma bundles both rank-256 LoRA adapters and the trained RMSNorm/modulation weight deltas in one ComfyUI-ready safetensors, letting users reproduce the original Krea 2 fine-tune behaviour without needing the full checkpoint.
What Sets It Apart
- LoRA + full-weight deltas: includes 264 adapter tensors plus ~159 RMSNorm/modulation .diff tensors, so loading at strength=1.0 reproduces the fine-tune more exactly than a "pure" LoRA that omits norm/mod deltas.
- ComfyUI-first packaging: single safetensors file formatted for ComfyUI's LoraLoader, simplifying application in Comfy pipelines and ensuring the extra .diff deltas are applied automatically.
- Tuned for Krea 2 family: produced from a Krea 2 fine-tune and compatible with both Krea-2 base and the few-step distilled Turbo checkpoint (Turbo recommended for faster sampling).
- Explicit provenance & license: conversion provenance included in the card and distributed under MIT; note the Krea 2 base weights are governed by their own license.
Who it's for — fit and tradeoffs
Great fit if you want reproducible fine-tune outputs on Krea 2 without storing or sharing the full model checkpoint — artists and ComfyUI users who need the exact trained normalization/modulation behavior will benefit. Look elsewhere if you need a tiny, latency-optimized adapter (Kroma is ~1.88 GB and includes full-weight deltas) or if you cannot comply with the base Krea 2 model's license terms. For inference, expect different recommended sampling regimes for base vs Turbo Krea 2 (Turbo: few-step distilled workflow; base: multi-step settings).
