Krea: Krea 2 Raw
krea-2-raw- Price
- From $0.015
- Modalities
- Image
About Krea 2 Raw
Krea 2 Raw is the base checkpoint of the Krea 2 family, released as an open-weight 12B text-to-image model before the post-training and distillation applied to Krea 2 Turbo. Krea describes it as an undistilled model for fine-tuning and post-training research. It suits LoRA training and custom domain work more than general-purpose generation.
Where it works well
- Open weights and an undistilled design make it a base for fine-tuning and LoRA training.
- It sits before the later post-training stages, so it keeps a wide visual range to specialize.
- A 12B Diffusion Transformer with a Qwen 3 VL text encoder gives a documented architecture to build on.
- Text-to-image and image-to-image are supported.
When to choose another model
- Krea says it fits custom work better than general inference; for ready-to-use images pick Turbo, Medium, or Large.
- Without distillation it needs more steps than Krea 2 Turbo, so each image takes longer.
- Krea's report calls the model undertrained and lists native 2K to 4K resolution as future work.
Getting started
Create API key
Create a key in Console, then use it with every model on the platform.
Send your first request
Copy the example for your language and run it against the endpoint.
Image editingText to imageTokenLab endpointPOST/v1/images/generationscurl -X POST "https://api.tokenlab.sh/v1/images/generations" \ -H "Authorization: Bearer sk-xxx" \ -H "Content-Type: application/json" \ -d '{ "operation": "image-to-image", "model": "krea-2-raw", "prompt": "A minimalist product photo of matte black headphones on a soft blue background.", "image_url": "https://example.com/source.png" }'
Parameters
Image editing
| Parameter | Values and limits | Default | Notes |
|---|---|---|---|
cfg_scale | ≥ 0, ≤ 30 | Not provided | Use multiples of 0.01. |
guidance_scale | ≥ 0, ≤ 30 | Not provided | Use multiples of 0.01. |
height | ≥ 64, ≤ 2048 | 1024 | Use an integer. Use multiples of 16. |
image | Not provided | Not provided | Alias of the combined images input. See input_limits.images for the total input image count. |
image_url | Not provided | Not provided | Alias of the combined images input. See input_limits.images for the total input image count. |
image_urls | Not provided | Not provided | Alias of the combined images input. See input_limits.images for the total input image count. |
images | Inputs: ≥ 1, ≤ 4 | Not provided | Limits apply when this input is supplied. Image aliases are combined into one input image list; the limit applies to the combined list. |
n | 1 | 1 | Use an integer. |
negative_prompt | Not provided | Not provided | Minimum length: 2 characters. Maximum length: 3000 characters. |
operationRequired | Not provided | Not provided | Not provided |
output_format | jpeg, png, webp | Not provided | Default encoding is shared by these equivalent values: jpg, jpeg. |
prompt | Not provided | Not provided | Minimum length: 2 characters. Maximum length: 10000 characters. |
seed | ≥ 0, ≤ 9007199254740991 | Not provided | Use an integer. |
steps | ≥ 1, ≤ 100 | 20 | Use an integer. |
width | ≥ 64, ≤ 2048 | 1024 | Use an integer. Use multiples of 16. |
Text to image
| Parameter | Values and limits | Default | Notes |
|---|---|---|---|
cfg_scale | ≥ 0, ≤ 30 | Not provided | Use multiples of 0.01. |
guidance_scale | ≥ 0, ≤ 30 | Not provided | Use multiples of 0.01. |
height | ≥ 64, ≤ 2048 | 1024 | Use an integer. Use multiples of 16. |
images | Not provided | Not provided | Text-to-image does not accept source images. Do not send this field for that operation. Use one of this model's declared image operations for source-image requests. |
n | 1 | 1 | Use an integer. |
negative_prompt | Not provided | Not provided | Minimum length: 2 characters. Maximum length: 3000 characters. |
operationRequired | Not provided | Not provided | Not provided |
output_format | jpeg, png, webp | Not provided | Default encoding is shared by these equivalent values: jpg, jpeg. |
prompt | Not provided | Not provided | Minimum length: 2 characters. Maximum length: 10000 characters. |
seed | ≥ 0, ≤ 9007199254740991 | Not provided | Use an integer. |
steps | ≥ 1, ≤ 100 | 20 | Use an integer. |
width | ≥ 64, ≤ 2048 | 1024 | Use an integer. Use multiples of 16. |
Fields outside this contract are rejected.
Pricing
The Verified price applies to Verified, which costs less on most models. The Official price is the model maker's published price and applies to the more reliable Official route. Auto bills the route that completes the request.
Image editing
per image- Official price
- $0.015
- Verified price
- —
- Discount
- —
Text to image
per image- Official price
- $0.015
- Verified price
- —
- Discount
- —
| Official priceper image | Verified priceper image | Discount | |
|---|---|---|---|
| Image editing | $0.015 | — | — |
| Text to image | $0.015 | — | — |
Usage & activity
Success rate is the share of requests that completed. Latency is how long a full response takes; P95 means 95% of requests finished within that time.
Usage & availability
Last 24 hours- Requests
- Success rate
- P95 latency
- Total tokens
Data is based on aggregate user requests, excluding status checks.
Open in Console
Open Krea 2 Raw in Console with a prompt ready to edit or send.
Help me create an image with krea-2-raw using image-edit at /v1/images/edits. Show the result, cost, and any limits I should know.
Use cases
Best for- Text to Image
LoRA training
Train a LoRA on your product, character, or house style, starting from the undistilled checkpoint.
Domain-specific models
Fine-tune for a narrow field such as architectural renderings or medical illustration.
Research baselines
Compare your own post-training or distillation methods against an open, undistilled reference model from a real production lab.
Wide-diversity sampling
Explore a broad range of aesthetics before choosing which directions to specialize.
Prompt examples
A technical cutaway illustration of a mechanical watch, clean lines, labeled parts without text.
Isometric room scene, wooden furniture, plants, afternoon light, matte colors.
A candid photo of a market vendor laughing, available light, natural skin texture.
FAQ
What is Krea 2 Raw?
The base release checkpoint of Krea 2, an undistilled open-weight model that Krea published for fine-tuning and post-training research. It comes before the distillation applied to Krea 2 Turbo.
Should I use Raw or Turbo for everyday images?
Turbo. Krea's own guidance is to train on Raw and generate with Turbo, because Raw is meant for customizing, such as LoRA work, rather than everyday generation.
What architecture does it use?
According to the technical report, it is a single-stream Diffusion Transformer with a Qwen 3 VL text encoder, scaled to 12 billion parameters and trained in stages.
Which license applies to it?
Krea describes a permissive license for weights and inference, and press coverage calls it a custom license. Read the license text itself before commercial use.
How much does Krea 2 Raw cost?
On TokenLab, Krea 2 Raw costs $0.015 per image. The pricing table above shows the full breakdown. Rates depend on the billing unit, specification, and usage. Compare matching conditions in the model's detailed pricing; a single rate does not determine the total cost.
Which endpoint should Krea 2 Raw use?
Use https://api.tokenlab.sh/v1/images/edits for Krea 2 Raw. The request example below shows the matching code shape.
Which operations does Krea 2 Raw support?
Krea 2 Raw supports Image editing, Text to image. Select an operation above to see its endpoint and request example.
Compare Krea 2 Raw
Sources
Reviewed Oct 2, 2026