The ComfyUI repackaging makes Krea 2's two-checkpoint workflow easy to run and iterate locally: train and tune on the undistilled RAW checkpoint, then run fast, few-step generation on the distilled Turbo checkpoint. That split lets creators develop LoRAs and style-specific tweaks on a malleable base while keeping inference cheap and quick.
What Sets It Apart
- RAW vs Turbo workflow: RAW is an undistilled, high-variation base suitable for fine-tuning and LoRA training; Turbo is distilled for 8-step sampling and consistent, high-quality outputs. This explicit pairing is designed so LoRAs trained on RAW express well on Turbo.
- ComfyUI packaging: model, text-encoder, VAE and several LoRAs are laid out for immediate use in ComfyUI, lowering the friction to run local experiments and iterate prompts or LoRAs.
- Modern stack: the open-release uses a Qwen Image VAE, a Qwen3‑VL text encoder with multi-layer feature aggregation and a DiT-based backbone, making it compatible with common local inference and fine-tuning toolchains (Diffusers/ComfyUI workflows).
Who It's For & Trade-offs
Great fit if you: want a local text-to-image stack that separates training and inference (train LoRAs on RAW, run on Turbo); need ComfyUI-ready checkpoints for experimentation; or prefer running models and LoRAs offline. Look elsewhere if you: require turnkey cloud-hosted APIs, need models with enterprise commercial licensing already purchased (this release uses a community license), or need a model tuned only for maximal diversity rather than distilled consistency.
Where It Fits
Use this package to prototype styles and LoRAs locally, then deploy distilled Turbo-based inference for fast iteration. For heavy production inference or commercial licensing, consult the upstream project's licensing and commercial options before deployment.
