Krea 2 Turbo in one paragraph: 8 steps, CFG 0, open weights
Krea 2 Turbo (written krea2 turbo or krea-2-turbo) is the distilled half of Krea's open-weight Krea 2 release from June 2026. It generates in 8 sampling steps instead of RAW's 52, runs at CFG 0.0 because guidance was distilled into the weights, outputs 1K to 2K, and ships under the Krea AI Community License. The FP8 build is about 13 GB and runs on a 12–16 GB card; the BF16 build is 26.3 GB and wants 24 GB or more. LoRAs trained on RAW load straight onto Turbo.
That is the short version. The rest of this page is the specs, the RAW-versus-Turbo comparison people actually search for, the ComfyUI settings that work, and what to do if you want the output without a 13 GB download.
Krea 2 Turbo specs and download sizes
| Spec | Krea 2 Turbo |
|---|---|
| Type | Timestep- and guidance-distilled checkpoint |
| Architecture | Diffusion transformer, single-stream blocks with SwiGLU MLPs and grouped-query attention |
| Text encoder | Qwen 3 VL (4B, FP8 scaled in the ComfyUI build) |
| VAE | Qwen Image VAE |
| Sampling steps | 8 |
| CFG | 0.0 as distilled; community guides use ~1.0 |
| Resolution | 1K to 2K (set 2.0 megapixels for 2K) |
| Released | June 2026 |
| License | Krea AI Community License |
The text encoder matters more than it looks. Qwen 3 VL takes both text and images, and Krea's technical report says it was chosen over T5-XXL for exactly that reason. It is also why Krea 2 restyles rather than edits — the image goes in as meaning, not as pixels.
Model files, from the ComfyUI build
| File | Size / target |
|---|---|
krea2_turbo_fp8_scaled.safetensors |
~13 GB, recommended for 12–16 GB VRAM |
| BF16 build | 26.3 GB, for 24 GB+ cards |
krea2_turbo_int8_convrot.safetensors |
Int8, used by the style-reference workflow |
qwen3vl_4b_fp8_scaled.safetensors |
Text encoder |
qwen_image_vae.safetensors |
VAE |
| NVFP4 / MXFP8 builds | Higher-end hardware |
On a 24 GB card the full stack sits around 18.6 GB resident once the diffusion model, text encoder and VAE are loaded.
Place them like this:
ComfyUI/
└── models/
├── diffusion_models/
│ └── krea2_turbo_fp8_scaled.safetensors
├── text_encoders/
│ └── qwen3vl_4b_fp8_scaled.safetensors
├── vae/
│ └── qwen_image_vae.safetensors
└── loras/
└── krea2_softwatercolor.safetensors
Krea 2 RAW vs Krea 2 Turbo: the comparison table
This is the single most searched question about the release, so here it is without the preamble.
| Krea 2 RAW | Krea 2 Turbo | |
|---|---|---|
| Distillation | None, base model | Timestep- and guidance-distilled |
| Sampling steps | 52 (full-step) | 8 |
| CFG | Conventional guidance, ~3.0 in edit workflows | 0.0 as distilled |
| Output character | Higher variation, more malleable, finer micro-texture | Consistent, polished, slightly smoothed |
| Built for | Fine-tuning and LoRA training | Production inference |
| Relative speed | Baseline | Roughly 8x faster in community benchmarks |
| LoRA compatibility | LoRAs are trained here | RAW-trained LoRAs load directly |
Speed figures from community testing vary a lot by hardware — reported ratios cluster around 8x, with absolute times ranging from a couple of seconds to a minute and a half per image depending on the card. Treat the ratio as the reliable number and the absolute times as anecdotes.
The verdict: download Turbo. RAW earns its 52 steps only if you are training a LoRA or fine-tuning, and the workflow Krea designed is explicitly "train on RAW, infer on Turbo". If you are generating images, RAW costs you six times the steps for a difference most people cannot see outside a 100% crop.
Krea 2 Turbo settings that work in ComfyUI
The single biggest mistake is bringing Stable Diffusion habits to the CFG field. Krea 2 Turbo is distilled at CFG 0.0 and will fall apart at 7–15.
steps: 8 (default; 12 if you are doing an img2img pass)
cfg: 0.0–1.0 (2.5 for img2img in community workflows)
resolution: ResolutionSelector, megapixels = 2.0 for 2K
lora strength: 1.0 (Krea's official style LoRAs)
prompt_enhance: on (LLM prompt expansion, on by default)
Getting it running from scratch
- Update ComfyUI to a recent build — the Krea-2 template and its subgraph nodes will not load on an older version.
- Drop
krea2_turbo_fp8_scaled.safetensorsintomodels/diffusion_models/,qwen3vl_4b_fp8_scaled.safetensorsintomodels/text_encoders/, andqwen_image_vae.safetensorsintomodels/vae/. - Search "Krea-2" in the Template Library and open the text-to-image workflow.
- Set the ResolutionSelector, type a prompt into the subgraph, and queue. The defaults — 8 steps, prompt enhancement on, no LoRA — already produce a usable image.
- Only then start changing things. Most "Krea 2 Turbo is broken" reports are someone who changed CFG first.
Krea shipped nine official style LoRAs, each with a trigger word and each documented at strength 1.0: darkbrush, dotmatrix, kidsdrawing, neondrip, rainywindow, retroanime, softwatercolor, sunsetblur and vintagetarot. Drop the .safetensors files into ComfyUI/models/loras/ and pick one in the CustomCombo node.
Prompting is the other adjustment. Krea 2 is an aesthetics-first model with an LLM expanding your prompt before it reaches the sampler, so short direction-setting prompts beat long keyword strings. The same instinct applies across modern models — the reasoning is laid out in the GPT Image 2 prompting guide.
Style reference: the closest thing Krea 2 Turbo has to img2img
Krea's first-party image-input feature is style reference, not img2img. In ComfyUI it is a separate template that pairs the Int8 Convrot build with a dedicated krea2_style_reference.safetensors LoRA and takes one or two reference images, each with its own strength.
It transfers palette, medium, lighting and mood. It does not preserve your subject. What that means in practice, which denoise ranges people actually use, and what to do when the face drifts is covered in full in Krea 2 img2img.
Running Krea 2 Turbo without downloading anything
To be clear: Img2Img AI does not serve Krea 2 Turbo. Krea's weights are gated and the model is text-to-image, which is a different product to what this site does.
What it does do is the job most people are chasing when they land on a Krea 2 Turbo page — take an image you already have and change it, in a browser, in seconds, on models chosen for keeping the subject intact. No 13 GB checkpoint, no VRAM floor, no node graph. If you specifically want Krea 2's aesthetic, Krea's own web app is the right place; if you want a restyled or edited version of your own photo today, start here.
For a fast open-weight text-to-image model you can compare against, Z-Image Turbo is the closest neighbour: lighter, quicker, better at legible text, less aesthetic range.
What people compare Krea 2 Turbo against
Krea 2 Turbo vs Krea 2 RAW. Covered above. Turbo unless you are training.
Krea 2 Turbo vs Z-Image Turbo. The two open-weight speed models people weigh against each other. Krea 2 Turbo wins on aesthetic range, style reference and the official style LoRAs. Z-Image Turbo is lighter to run and noticeably better at keeping small text legible. Prompting habits do not transfer between them — both reject the high-CFG keyword-stuffing style, but in different ways.
Krea 2 Turbo vs Krea's hosted models. Turbo is the fast tier. Krea's app also exposes larger Krea 2 variants aimed at photorealism, at a higher cost per image. If you are already running locally, the open Turbo checkpoint is the one you have.
Krea 2 Turbo vs an editing model. Not a real comparison — Turbo generates, editors preserve. That distinction is the whole of can Krea 2 edit images.
Where Krea 2 Turbo falls short
- It cannot edit. There is no inpainting, no outpainting, no region locking. The community Identity Edit LoRA is the only route to anything edit-like, and it needs a custom node pack — see can Krea 2 edit images.
- 2MP ceiling. Native output tops out around 2K. Anything larger is an upscale.
- Text in images. Distilled 8-step sampling is not kind to small type.
- Gated weights. The Hugging Face repo requires accepting the licence, which breaks unattended pulls.
- Smoothing. Distillation trades micro-texture for speed. At 100% zoom RAW holds detail that Turbo washes out.
Quick answers about Krea 2 Turbo
What is the difference between Krea 2 and Krea 2 Turbo?
Krea 2 is the model family; RAW and Turbo are its two open checkpoints. RAW is the undistilled base at 52 steps, built for fine-tuning and LoRA training. Turbo is the distilled 8-step checkpoint built for generating. LoRAs trained on RAW run on Turbo.
What settings should I use for Krea 2 Turbo?
8 steps, CFG 0.0 to 1.0, ResolutionSelector at 2.0 megapixels for 2K, style LoRAs at strength 1.0. For an img2img pass, community workflows land around 12 steps at CFG 2.5.
How do I prompt Krea 2 Turbo?
Short, direction-led prompts. An LLM expands your prompt before sampling, so listing keywords works against you. Describe the look and the subject, then let the style LoRA or style reference carry the medium.
How many parameters does Krea 2 have?
Krea's technical report does not publish a parameter count. Community documentation for the RAW checkpoint reports 12.9B.
How much VRAM does Krea 2 Turbo need?
The FP8 build at roughly 13 GB runs on 12–16 GB cards. BF16 at 26.3 GB wants 24 GB or more. On a 24 GB card the loaded stack sits near 18.6 GB.
Can Krea 2 Turbo edit an existing photo?
Not natively. It will regenerate the photo in a new style rather than edit it. For actual edits, either add the community Krea2Edit node pack or use a browser editor built for the job.
Is Krea 2 Turbo free?
The weights are free to download under the Krea AI Community License, so local inference costs you only hardware and electricity. Running it through Krea's hosted app or a third-party API is paid.
Get the result without the download
If what you actually want is a changed version of an image you already have, the checkpoint is a detour. Upload it to Img2Img AI, describe the change, name what must stay, and get it back in seconds.
