Krea 2New

Krea 2 is Krea AI's first foundation image model, trained from scratch to prioritize how an image feels — its mood, lighting, and texture — rather than just what it contains. It's served on Civitai through the official fal.ai API partnership, so you can generate with the closed-weights hosted build directly on-site with no GPU or install. The model is designed around aesthetic control: style references with per-reference strength, moodboards, and a tunable creativity dial (raw, low, medium, high) that governs how far the model may drift from your inputs, making the prompt only one input among several.

The hosted model ships in two sizes. Large is the bigger build, with softer post-training that leaves more of the raw base character intact; it's the pick for photorealism of humans and animals, motion blur, film grain, low dynamic range, and imperfect lighting that reads as a real photograph. Medium is smaller, faster, and cheaper, with heavier post-training that pushes it toward illustration, anime, painting, and other stylized aesthetics with clean line work. Alongside these, Krea released two open-weight checkpoints on Hugging Face: Raw, the undistilled full-guidance build (Krea recommends ~52 steps at CFG 3.5) that carries the highest quality ceiling and is the intended target for fine-tuning and LoRA training; and Turbo, an 8-step distilled build of Raw that runs without classifier-free guidance for fast inference at 1K–2K resolution (up to 2048×2048).

Choose Krea 2 when you want photoreal grit or a defined illustration style with minimal fuss — it has a noticeable edge on the hard cases other models fight you on: lens flares, chrome and metallic surfaces, motion blur, glitter and iridescent textures, film grain, and starburst highlights. Reach for Large for chrome, lens flares, and photographic realism; Medium for clean stylized art; Raw when you want the maximum quality ceiling or a predictable base for training; and Turbo when you want Krea 2 aesthetics at a fraction of the compute. LoRAs trained on Raw carry over to Turbo, so you can train once on the full model and run inference fast on the distilled build.