Krea 2 RAW & Turbo Hit ComfyUI — Let's Go 🔥

🔥 BREAKING: Krea 2 — the image generation model that's been living rent-free in every AI artist's head — just dropped its open-source weights directly into ComfyUI. And no, this isn't a drill. Two whole checkpoints. For free. On your own hardware. You can stop screaming now (but honestly, we wouldn't blame you).

Krea shipped Krea 2 RAW and Krea 2 Turbo on June 24, 2026, and they load through ComfyUI's standard diffusion model nodes like they've always been there. No proprietary BS, no vendor lock-in, no "you need to use our cloud" nonsense. Just drop the weights in and let your GPU do the talking.

This is the kind of move that makes the open-source AI community stand up and slow clap. Let's break down why this actually matters beyond the hype.

RAW vs Turbo: Pick Your Fighter 🎮

Krea didn't just dump one checkpoint and call it a day. They shipped two models designed to work together like PB&J:

Krea 2 RAW — This is the base model. Untouched, undistilled, no post-training shenanigans. It's like getting the dough before it's baked. RAW is diverse, malleable, and perfect for fine-tuning, LoRA training, and research. It's the creative chaos you can shape into anything.
Krea 2 Turbo — The 8-step distilled checkpoint built for speed. Think of it as the espresso shot version — fast, high-quality, and ready to serve. It's distilled from a fully post-trained RL checkpoint, so it produces cleaner outputs but with less variation between generations.

Here's the secret sauce: you train a LoRA on RAW, then apply it on Turbo. The LoRA expresses strongly on Turbo, giving you the best of both worlds — the malleability of RAW for training and the speed of Turbo for inference. That's not just a feature, that's a whole vibe.

Peeking Under the Hood 🧠

For the nerds in the back (we see you, and we respect you): Krea 2 OSS is a 12B dense DiT (Diffusion Transformer). It uses the Qwen Image VAE and a Qwen3-VL text encoder with multi-layer feature aggregation. Translation? It's a beefy model that understands complex prompts and produces consistently impressive results across a wide aesthetic range — from photographic portraits to anime to high-contrast editorial work.

Speed-wise, Turbo generates in just 8 steps. You'll be iterating faster than you can say "just one more tweak." (Spoiler: there is never "just one more tweak.")

"Krea 2 OSS covers a wide aesthetic and stylistic range, from photographic portraits to anime to high-contrast editorial macro work." — ComfyUI Newsletter

Why This Matters (Read: Why You Should Care)

Open-weight models landing in ComfyUI is a big deal for three reasons:

Local generation is king. No API bills, no rate limits, no "sorry, our servers are overloaded." You own your pipeline. Your GPU. Your rules.
The RAW + Turbo pairing is smart design. Most open models make you choose between flexibility and speed. Krea 2 gives you both in one package.
Zero lock-in. It loads through standard ComfyUI nodes. If tomorrow you want to swap it out for another model, you can. The workflow stays. The ecosystem wins.

This is the kind of release that makes you want to rebuild your entire workflow from scratch just to try it. And honestly? You probably should.

The Verdict

Krea 2 going open-source in ComfyUI is a massive W for the local AI art community. Whether you're a LoRA trainer, a workflow architect, or just someone who wants to generate cool stuff without signing up for another SaaS subscription — this release has your name written all over it.

Go download the checkpoints, fire up ComfyUI, and see what the RAW + Turbo combo can do. Your next masterpiece is waiting. And it's running entirely on your machine.

— Node wrangling, one checkpoint at a time