Krea's own foundation image model, trained from scratch to avoid the plastic AI look, in its fast distilled Turbo form. 1 credit per image, five aspect ratios, no watermark, on every Aura AI plan.
Krea 2 is the first foundation image model that Krea AI built from scratch rather than fine-tuning someone else's checkpoint. Krea announced it on May 12, 2026 and open-sourced it on June 23, 2026 as a 12.9 billion parameter diffusion transformer with a permissive license. It follows Krea 1, the company's June 2025 debut model, which made its name by attacking the "AI look": the over-smooth skin, exaggerated lighting and formulaic compositions that make generated pictures easy to spot. Krea 2 keeps that aesthetic mission and adds a new architecture, a style-reference system and open weights.
The open release ships as two checkpoints. Krea 2 RAW is the undistilled base, diverse and malleable, meant for fine-tuning, post-training and LoRA work. Krea 2 Turbo is the guidance and timestep distilled version, built for fast few-step text-to-image inference, which Krea quotes at roughly two seconds per image on its own hardware. Under the hood, the technical report describes a single-stream diffusion transformer with grouped-query attention, sigmoid-gated attention, SwiGLU feed-forward layers and a Qwen 3-VL text encoder, trained progressively at 256, 512 and 1024 pixels.
On Aura AI, the Krea 2 Turbo checkpoint runs in text-to-image mode for 1 credit per image, in five aspect ratios, on every plan. It shares the 1 credit tier with FLUX 2 Turbo, Gemini 3.1 Flash and Hunyuan Image 3, and the difference between them is mostly taste: Krea 2 leans toward natural textures, film-like light and painterly styles rather than glossy render. Output is delivered without watermark and can go straight into the image editor or image-to-video.
Six reasons Krea's second model is worth a credit
Krea's stated goal since Krea 1 is images that do not read as generated: real skin texture, imperfect light, believable grain. Krea 2 continues that direction with a model trained from scratch on that preference.
Turbo is the guidance and timestep distilled checkpoint of Krea 2, designed for few-step inference. Upstream Krea quotes about two seconds per image, which is why it sits at the 1 credit tier on Aura AI.
A single-stream DiT with grouped-query attention, sigmoid-gated attention and SwiGLU layers, paired with a Qwen 3-VL text encoder that reads long, natural-language prompts rather than keyword lists.
Krea 2 was trained for diversity across photography, illustration, painting and graphic styles. Ask for a specific film stock, a watercolor wash or a risograph print and the result follows the medium, not a default render style.
The Krea 2 architecture includes a style-reference system that guides output with one or more reference images, each with adjustable strength. On Aura AI the model runs in plain text-to-image mode, so describe the style in words.
Both RAW and Turbo checkpoints are public on Hugging Face under a permissive license. Train a LoRA on RAW, apply it on Turbo, or simply use the hosted Turbo model on Aura AI without any setup.
Four steps from prompt to a downloadable image
A single flat cost, the same in every aspect ratio
| Variant | Mode | Credits per image | Best for |
|---|---|---|---|
| Krea 2 Turbo | Text to image | 1 credit | Natural-looking photos, illustration styles, fast iteration |
The four 1 credit image models on Aura AI, side by side
| Feature | Krea 2 Turbo | FLUX 2 Turbo | Gemini 3.1 Flash | Hunyuan Image 3 |
|---|---|---|---|---|
| Maker | Krea AI | Black Forest Labs | Tencent | |
| Text-to-image credits | 1 | 1 | 1 | 1 |
| Edit variant on Aura AI | No | FLUX 2 Edit, 2 credits | ✓ 1 credit | No |
| Open weights | ✓ Permissive license | Dev checkpoint only | No | ✓ Community License |
| Parameters | 12.9B | Not disclosed for Turbo | Not disclosed | 80B MoE (13B active) |
| Strength | Natural aesthetics, styles | Speed, photoreal | Speed, prompt following | Text rendering, dense scenes |
| Aspect ratios on Aura AI | 5 | 5 | 5 | 5 |
| Plan | All plans | All plans | All plans | All plans |
Where natural textures and a 1 credit price pay off
Social and editorial images that should pass as a real photo: imperfect light, real skin, a bit of motion blur. Krea 2 was tuned for exactly this kind of restraint.
Risograph, gouache, ink wash, screen print. Name the medium and the model follows its physical traits, including paper texture and ink bleed.
9:16 imagery for stories and reels that does not look like stock art. Generate several at 1 credit each, pick the best and animate it with an image-to-video model.
Explore a visual direction with ten quick generations before committing credits to a heavier model. Krea 2 Turbo is fast enough to keep the ideas flowing.
Common questions about Krea 2 Turbo on Aura AI
Natural-looking images from Krea's own foundation model, 1 credit each, five aspect ratios, no watermark. Available on every plan, worldwide.