Skip to main content

Krea.aiImage model

Krea 2 Raw

GraduallyBench Image v2not tested yet

n/aour own benchmark0 to 100

Arena ratingnot rated

n/aUser voteshigher is better

Priceno price

n/aper imagelower is better

PopularityRank 14 of 48 image models

49out of 100new calculation

based on two data groups, mostly search, range 39 to 58

Compare with another model

Summary

What is distinctive about Krea 2 Raw is its status as the 12-billion-parameter base checkpoint. Krea released it on June 22, 2026 before additional post-training and fine-tuning, as a research and adaptation point rather than a finished product model.

The download is gated under the Krea 2 Community License. Krea does not recommend Raw for regular inference and positions it as a base for custom adaptation and LoRAs; Krea 2 Turbo is the separately post-trained and distilled checkpoint.

All benchmark results

Every published result with its source, test setup and comparison models.

Show all 10 published benchmark results

Measurements without matching peer values

These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.

ImageBench (Show measurement, test conditions, and source)
Source value
61.5
Metric
Capability pass rate (0-100)
Unit
%
Benchmark version
V1.2 scores.csv at Hugging Face commit 9671ce9e2d0cc91099348135cbffcaa03159a00f
Category
text-to-image
Direction
Higher is better
Sample
192
Source type
independent-evaluator
Evaluator
ImageBench V1.2 (dh7)
Status
active
Retrieved at
2026-10-16
Methodology
Independent evaluation. ImageBench row label local/krea-2-raw; run target gx10/krea2-raw; 1024x1024, 52 inference steps, guidance_scale 3.5. The dataset has 192 generated images, one for each of 64 tests x 3 variants.
Context
The dataset identifies the numerator as 118 PASS results out of 192 generations. VLM verdicts are machine-assisted and may be wrong; ImageBench notes prompt-selection bias. The page family label is V1, while this exact value is sourced from the pinned V1.2 snapshot. ImageBench does not publish a weight hash, runtime precision, or quantization for the gx10 target; byte identity to the vendor files is unverified.
Source ImageBench V1.2 (dh7)
ImageBench category pass rate (Show measurement, test conditions, and source)
Source value
66.7
Metric
Graphical design pass rate
Unit
%
Benchmark version
V1.2 scores.csv at Hugging Face commit 9671ce9e2d0cc91099348135cbffcaa03159a00f
Category
text-to-image
Direction
Higher is better
Source type
independent-evaluator
Evaluator
ImageBench V1.2 (dh7)
Status
active
Retrieved at
2026-10-16
Methodology
Independent evaluation. ImageBench row label local/krea-2-raw, text-to-image. Source page specifies 1024x1024, 52 inference steps, guidance_scale 3.5. Category-specific denominator is not published in scores.csv.
Context
VLM PASS/FAIL labels may be wrong, and category-specific N is not published. Do not infer a category denominator from the total 192 generations. ImageBench does not publish a weight hash, runtime precision, or quantization for the gx10 target; byte identity to the vendor files is unverified.
Source ImageBench V1.2 (dh7)
ImageBench category pass rate (Show measurement, test conditions, and source)
Source value
33.3
Metric
Human realism pass rate
Unit
%
Benchmark version
V1.2 scores.csv at Hugging Face commit 9671ce9e2d0cc91099348135cbffcaa03159a00f
Category
text-to-image
Direction
Higher is better
Source type
independent-evaluator
Evaluator
ImageBench V1.2 (dh7)
Status
active
Retrieved at
2026-10-16
Methodology
Independent evaluation. ImageBench row label local/krea-2-raw, text-to-image. Source page specifies 1024x1024, 52 inference steps, guidance_scale 3.5. Category-specific denominator is not published in scores.csv.
Context
VLM PASS/FAIL labels may be wrong, and category-specific N is not published. Do not infer a category denominator from the total 192 generations. ImageBench does not publish a weight hash, runtime precision, or quantization for the gx10 target; byte identity to the vendor files is unverified.
Source ImageBench V1.2 (dh7)
ImageBench category pass rate (Show measurement, test conditions, and source)
Source value
81.5
Metric
Professional Studio pass rate
Unit
%
Benchmark version
V1.2 scores.csv at Hugging Face commit 9671ce9e2d0cc91099348135cbffcaa03159a00f
Category
text-to-image
Direction
Higher is better
Source type
independent-evaluator
Evaluator
ImageBench V1.2 (dh7)
Status
active
Retrieved at
2026-10-16
Methodology
Independent evaluation. ImageBench row label local/krea-2-raw, text-to-image. Source page specifies 1024x1024, 52 inference steps, guidance_scale 3.5. Category-specific denominator is not published in scores.csv.
Context
VLM PASS/FAIL labels may be wrong, and category-specific N is not published. Do not infer a category denominator from the total 192 generations. ImageBench does not publish a weight hash, runtime precision, or quantization for the gx10 target; byte identity to the vendor files is unverified.
Source ImageBench V1.2 (dh7)
ImageBench category pass rate (Show measurement, test conditions, and source)
Source value
75.4
Metric
Spatial Reasoning pass rate
Unit
%
Benchmark version
V1.2 scores.csv at Hugging Face commit 9671ce9e2d0cc91099348135cbffcaa03159a00f
Category
text-to-image
Direction
Higher is better
Source type
independent-evaluator
Evaluator
ImageBench V1.2 (dh7)
Status
active
Retrieved at
2026-10-16
Methodology
Independent evaluation. ImageBench row label local/krea-2-raw, text-to-image. Source page specifies 1024x1024, 52 inference steps, guidance_scale 3.5. Category-specific denominator is not published in scores.csv.
Context
VLM PASS/FAIL labels may be wrong, and category-specific N is not published. Do not infer a category denominator from the total 192 generations. ImageBench does not publish a weight hash, runtime precision, or quantization for the gx10 target; byte identity to the vendor files is unverified.
Source ImageBench V1.2 (dh7)
ImageBench category pass rate (Show measurement, test conditions, and source)
Source value
73.3
Metric
Text Rendering pass rate
Unit
%
Benchmark version
V1.2 scores.csv at Hugging Face commit 9671ce9e2d0cc91099348135cbffcaa03159a00f
Category
text-to-image
Direction
Higher is better
Source type
independent-evaluator
Evaluator
ImageBench V1.2 (dh7)
Status
active
Retrieved at
2026-10-16
Methodology
Independent evaluation. ImageBench row label local/krea-2-raw, text-to-image. Source page specifies 1024x1024, 52 inference steps, guidance_scale 3.5. Category-specific denominator is not published in scores.csv.
Context
VLM PASS/FAIL labels may be wrong, and category-specific N is not published. Do not infer a category denominator from the total 192 generations. ImageBench does not publish a weight hash, runtime precision, or quantization for the gx10 target; byte identity to the vendor files is unverified.
Source ImageBench V1.2 (dh7)
ImageBench category pass rate (Show measurement, test conditions, and source)
Source value
44.4
Metric
Truthfulness pass rate
Unit
%
Benchmark version
V1.2 scores.csv at Hugging Face commit 9671ce9e2d0cc91099348135cbffcaa03159a00f
Category
text-to-image
Direction
Higher is better
Source type
independent-evaluator
Evaluator
ImageBench V1.2 (dh7)
Status
active
Retrieved at
2026-10-16
Methodology
Independent evaluation. ImageBench row label local/krea-2-raw, text-to-image. Source page specifies 1024x1024, 52 inference steps, guidance_scale 3.5. Category-specific denominator is not published in scores.csv.
Context
VLM PASS/FAIL labels may be wrong, and category-specific N is not published. Do not infer a category denominator from the total 192 generations. ImageBench does not publish a weight hash, runtime precision, or quantization for the gx10 target; byte identity to the vendor files is unverified.
Source ImageBench V1.2 (dh7)
ImageBench (Show measurement, test conditions, and source)
Source value
44.2
Metric
Overall composite (capability and estimated preference; 0-100 scale)
Unit
points
Benchmark version
V1.2 scores.csv at Hugging Face commit 9671ce9e2d0cc91099348135cbffcaa03159a00f
Category
text-to-image
Direction
Higher is better
Sample
192
Source type
independent-evaluator
Evaluator
ImageBench V1.2 (dh7)
Status
active
Retrieved at
2026-10-16
Methodology
Independent evaluation. ImageBench row label local/krea-2-raw, which the official Krea model card identifies by name as its Raw base-release checkpoint; this name-level mapping does not prove which weights gx10 loaded. ImageBench model page gives run target gx10/krea2-raw, 1024x1024, 52 inference steps, guidance_scale 3.5, and negative prompt 'blurry, low quality, distorted, deformed, low resolution'. The pinned V1.2 dataset defines 64 tests x 3 variants and one generation per prompt.
Context
The model page labels the family V1 while the exact row is in the pinned V1.2 dataset snapshot; the score is taken from that V1.2 row, not inferred from the rounded page title. ImageBench warns that VLM PASS/FAIL judgments can be wrong and that its prompt set favors tests that separate models. Inspect the published images; this score is not directly comparable to unrelated protocols. ImageBench does not publish a weight hash, runtime precision, or quantization for the gx10 target; byte identity to the vendor files is unverified.
Source ImageBench V1.2 (dh7)
ImageBench (Show measurement, test conditions, and source)
Source value
26.9
Metric
Estimated Preference Score (HPSv3-based; 0-100)
Unit
points
Benchmark version
V1.2 scores.csv at Hugging Face commit 9671ce9e2d0cc91099348135cbffcaa03159a00f
Category
text-to-image
Direction
Higher is better
Sample
164
Source type
independent-evaluator
Evaluator
ImageBench V1.2 (dh7)
Status
active
Retrieved at
2026-10-16
Methodology
Independent evaluation. The pinned V1.2 CSV reports EPS=26.9. The model page embedded result payload reports eps_n=164. The publisher does not explain how this EPS denominator is constructed; it is not equated with the 192 generated images.
Context
The page exposes eps_n=164, but does not document which inputs this denominator counts. It is not the total 192 generated images. ImageBench warns that VLM verdicts can be wrong and its prompt set favors separating models. The page labels the family V1 while the score is from the pinned V1.2 snapshot.
Source ImageBench V1.2 (dh7)
ImageBench generation latency (Show measurement, test conditions, and source)
Source value
912.2
Metric
Mean generation latency per image
Unit
seconds
Benchmark version
ImageBench V1 model page; same exact local checkpoint as pinned V1.2 scores.csv
Category
text-to-image
Direction
Lower is better
Sample
192
Source type
independent-evaluator
Evaluator
ImageBench V1 model page
Status
active
Retrieved at
2026-10-16
Methodology
Independent evaluation. Local run target gx10/krea2-raw; 1024x1024; 52 inference steps; guidance_scale 3.5. The page reports mean, minimum and maximum latency for 192 generated images.
Context
The page does not identify the underlying hardware or a latency measurement procedure. This local-run average is not comparable to API-provider latency or other hardware and is kept separate from the V1.2 quality scores. ImageBench does not publish a weight hash, runtime precision, or quantization for the gx10 target; byte identity to the vendor files is unverified.
Source ImageBench V1 model page

Popularity compared

How often the model is searched for, used and mentioned, compared with every model of its category.

Popularity index

0 to 100, higher is more popular

20 of 75 models

As of November 2026 · Source: gradually.ai popularity index

Over time

Weekly values from October 18, 2026 to November 20, 2026

Krea 2 RawSeedream 5.0 Lite 49Stable Diffusion 3.5 Large 48Krea 3 48MOSS-Image-Base 48

Method changes in the timeline, a jump at these points is no market event:

  1. 1November 1, 2026: Value only with search or usage data, 30-day attention
  2. 2November 1, 2026: Method version 4 (common scale, shrinkage)
  3. 3November 9, 2026: New topics measurement V20 (daily quorum, short forms)

What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.

All specifications

Architecture

Parameters
12B
Source: Parameters
Model class
Text-to-image diffusion model
Source: Model class

Watermarking and provenance

AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.

AI watermarking
Unknown
no source
Watermarking technique
Not known
no source

Licence and access

Weights
Open weights
Source: Weights
Access
Local
Source: Access
License
Krea 2 Community License
Source: License
Shutdown
None announced
no source

API

Model ID
krea/Krea-2-Raw
Source: Model ID

Further details

Checkpoint
krea/Krea-2-Raw, an undistilled research starting point before further post-training and fine-tuning.
Source: Checkpoint
Research role
The model card does not recommend the Raw checkpoint for regular inference, but as a base for fine-tuning and post-training.
Source: Research role
Access and license
Gated weight access after agreeing to the Krea 2 Community License.
Source: Access and license
Checkpoint role
Base checkpoint before additional post-training and fine-tuning
Source: Checkpoint role
Inference note
Not recommended for regular inference, intended as a base for fine-tuning and post-training
Source: Inference note

Sources and data date

Show all 12 sources

Benchmark sources at a glance

Which independent evaluators tested this model. This is coverage, not a quality ranking.

No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.

Without vendor claims and historical or disputed results; those stay labelled in the single values.

TypeEvidence and data date
Research statusResearch date October 4, 2026. 8 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication.
Not found in the inspected sources

raw-checkpoint-benchmark

Original wording

Krea 2 Raw precedes post-training and is not a substitute for the separately evaluated Large, Medium, or Medium Turbo serving checkpoints.

Not found in the inspected sources

krea-2-raw-independent-image-score

Note

Neither checked independent text-to-image leaderboard lists the exact Raw checkpoint. Scores for Krea 2 Turbo or other variants are not substitutes.

Not published

krea-2-raw-vendor-benchmark

Note

The official source does not publish a quantitative Raw-specific benchmark.

Not found in the inspected sources

krea-2-raw-independent-image-score

Note

No exact Raw result was found in the current image leaderboards. Results for Krea 2's separate distilled checkpoints are not assigned to Raw.

Not published

krea-2-raw-vendor-benchmark

Note

The official release page distinguishes Raw from distilled checkpoints but publishes no quantitative Raw quality benchmark.

Not found in the inspected sources

krea-2-raw-exact-image-benchmark

Note

No exact Raw checkpoint score was found in the checked current image leaderboards or official release page. Scores for separately distilled Krea 2 variants are not transferred.

Not published

krea-2-raw-vendor-quality-benchmark

Note

The current official Krea release/model materials checked in the preceding source audit distinguish Raw from distilled variants but publish no quantitative vendor quality result. This does not negate the independent ImageBench V1.2 results recorded here.

Additional sourceKrea.ai (retrieved October 3, 2026; October 4, 2026; October 16, 2026; November 1, 2026) · Krea 2 Raw model card · Editorial description reviewed October 3, 2026
Additional sourceImageBench V1.2 (dh7) (retrieved October 16, 2026)
Additional sourceImageBench V1 model page (retrieved October 16, 2026)
Additional sourceModel metadata (retrieved October 15, 2026; October 16, 2026)
Additional sourceModel metadata (retrieved October 16, 2026)12
Additional sourceOpenRouter: Popularity index (derived), CC BY 4.0
Additional sourceVercel AI Gateway: Popularity index (derived), CC BY 4.0
Additional sourceOllama: cumulative downloads for the popularity index
Additional sourceReplicate: cumulative runs for the popularity index
Additional sourceGoogle Trending Now: limited feed selection for Germany and the US
Additional sourceCreative Commons Attribution 4.0