krea2 AKSIOM Core Turbo v1

A balanced foundation model for realistic and semi-realistic image generation, developed as the core reference of the AKSIOM model family.

krea2_aksiom_core_turbo_v1.safetensors

Overview

krea2 AKSIOM Core Turbo v1 is the first Core release of the AKSIOM model family.

The model was originally developed as a foundation and development reference for exploring how different model components, LoRAs and related adaptations affect image generation. During development, it became a stable and useful model in its own right, and is now being released to the community as the starting point for the AKSIOM family.

The underlying foundation inherits its basic weighting from Krea-2 RAW, with additional development focused on:

• natural and neutral skin structures

• dynamic character formation

• realistic image generation

• stable semi-realistic generation

• analog photographic characteristics

• detailed and dynamic environments

• stronger interaction between subjects, backgrounds and overall scene composition

The goal was not to push the model toward a single fixed visual style. Instead, AKSIOM Core was developed to remain relatively balanced and adaptable, providing a useful base for different prompts, LoRAs and future model variations.

What is AKSIOM?

AKSIOM is planned as a growing model family rather than a single Model.

The Core models form the balanced foundation of the family. Future releases are intended to explore more specialized base models as well as additional Turbo variants, each focusing on different aspects of image generation while remaining connected to the AKSIOM foundation.

The idea is simple:

Core provides the foundation. Specialized models explore specific directions.

This release is therefore best understood as the first reference point of the AKSIOM family.

Key Characteristics

Realism & Semi-Realism

The model has been developed with a strong focus on realistic image formation while maintaining stability when working with semi-realistic concepts.

It is intended to handle transitions between realistic and stylized subjects without forcing every generation toward a purely photorealistic appearance.

Skin & Character Structure

Additional weighting was introduced to improve neutral skin structures and dynamic character formation.

The intention was to preserve natural-looking surfaces while allowing characters to remain responsive to different prompts, poses, environments and visual contexts.

Analog Photographic Characteristics

The model contains additional weighting aimed at reproducing characteristics associated with analog photography.

This is not intended as a strict photographic simulation. Instead, it provides another visual direction that can become apparent when prompts, composition and other model components support it.

Dynamic Backgrounds

Particular attention was given to detailed environments and backgrounds.

The goal was not simply to increase background detail, but to encourage backgrounds to participate in the overall composition and respond to the main subject and scene concept.

This can help produce environments that feel more connected to the generated scene rather than functioning purely as static detail behind the subject.

Composition & Adaptability

During development, the model was repeatedly evaluated with controlled generations covering different colors, poses, compositions, subjects and scene structures.

This helped maintain a relatively balanced response across different generation scenarios rather than optimizing the model around a single preferred composition or subject type.

Development Reference

One of the original purposes of Core Turbo was very different from a conventional production checkpoint.

The model served as a neutral development reference for the AKSIOM project.

During development, controlled generations were compared using histograms, graphs and repeated test configurations. These evaluations were used to observe how changes in model weights, LoRAs and other components affected:

• color distribution

• pose behavior

• composition

• character formation

• realism

• image detail

• background behavior

• model adaptability

• LoRA interaction and responsiveness

The purpose of this process was not to make Core Turbo visually neutral in the sense of producing bland images.

Rather, the goal was to maintain a sufficiently balanced reference model so that changes introduced by new components could be observed and compared more reliably.

In other words, Core Turbo was used as a measuring point during development, not simply as a final artistic target.

This role also influenced the design philosophy of the AKSIOM family.

Intended Use

AKSIOM Core Turbo v1 can be used for:

• realistic image generation

• semi-realistic image generation

• character-focused scenes

• environmental and cinematic compositions

• experimentation with LoRAs

• testing different prompting approaches

• evaluating visual adaptations

• developing derivative models

• exploring new AKSIOM family components

It can also be used as a general-purpose base for users who prefer a balanced model rather than a Unet|DiT strongly optimized toward one specific visual style.

Development Philosophy

AKSIOM Core was developed around a simple principle:

A strong base model should leave room for development.

Instead of attempting to maximize one particular visual characteristic, development focused on maintaining a useful balance between realism, detail, composition, character behavior and adaptability.

This makes the model particularly useful as a starting point for further experimentation.

The Core release is therefore not intended to represent the final state of the AKSIOM project. It is the foundation from which later models can evolve.

Model Family

The AKSIOM family is planned to contain several model directions.

AKSIOM Core

The balanced foundation of the family.

Designed to provide a broad and adaptable generation baseline.

AKSIOM Specialized Models

Future models may focus more strongly on specific generation characteristics or visual domains.

These models will build upon the experience and evaluation methods established during Core development.

AKSIOM Turbo

Turbo variants are intended for faster generation while retaining the characteristic behavior of the corresponding AKSIOM model direction.

Additional Turbo releases may therefore differ in focus from Core Turbo while remaining part of the same model family.

Recommended Use

The model is intentionally flexible and can be explored with a wide range of prompting styles.

For realistic results, prompts describing:

• subject

• environment

• lighting

• materials

• atmosphere

• composition

• action or pose

tend to provide useful control.

The model can also respond well to more conceptual or descriptive prompts when used with compatible workflows.

Because this is a Turbo model, experimentation with different sampling configurations is encouraged rather than assuming that one fixed setting represents the model's full capability.

LoRA Compatibility

One of the reasons for maintaining a balanced Core model was to provide a useful foundation for testing and developing LoRAs.

The model can therefore be used as a reference when evaluating how a LoRA changes:

• style

• color behavior

• character characteristics

• composition

• detail

• realism

• scene structure

Results will naturally depend on the architecture and training method of the individual LoRA.

The Core model should not be considered a guarantee that every LoRA will behave identically to its behavior on another Models.

Limitations

AKSIOM Core Turbo v1 is an experimental community release and should be considered part of an ongoing model-development project.

Generation quality can vary depending on:

• prompt structure

• sampling configuration

• resolution

• LoRAs

• workflow

• conditioning

• subject complexity

• scene composition

The development and evaluation process described above represents comparative testing within the AKSIOM development workflow. It should not be interpreted as a formal benchmark or scientific validation.

The model may also produce unintended anatomical, structural or compositional results, particularly with complex scenes or unusual prompts.

Background

This model originated from an internal development process focused on understanding and measuring model behavior rather than producing a single commercial-style Model.

Over time, the Core model became stable enough to serve as a useful standalone generation model.

Rather than keeping the checkpoint exclusively as an internal development tool, it is being released so that the community can experiment with it, build upon it and provide feedback that may help shape future AKSIOM releases.

Version

Model: krea2 AKSIOM Core Turbo v1
File: krea2_aksiom_core_turbo_v1.safetensors
Family: AKSIOM
Type: Core / Turbo
Base foundation: Krea-2 RAW
Status: Community Release
Version: v1

Credits & Base Model

This model inherits its fundamental weighting from Krea-2 RAW and contains additional development work and weight modifications.

Please refer to the applicable license and usage conditions of the underlying model before redistributing or creating derivative works.

License & Derivative Work

krea2 AKSIOM Core Turbo v1 is a derivative work based on the Krea-2 architecture and underlying Krea-2 RAW model. Its development and distribution are therefore subject to the applicable terms, restrictions and requirements of the Krea-2 Community License.

Users downloading, using, modifying or redistributing this model are responsible for reviewing and complying with the Krea-2 Community License and its applicable guidelines. This includes any requirements concerning attribution, derivative works, redistribution and permitted use.

The additional weights and development work introduced for AKSIOM Core Turbo do not replace, override or supersede the terms of the underlying Krea-2 Community License.

Base model: Krea-2 RAW
Derivative: krea2 AKSIOM Core Turbo v1

Final Notes

This release is intentionally called Core.

It is not meant to be the most specialized, most stylized or most extreme model in the AKSIOM family.

It is the model from which the family grows.

The development process behind Core Turbo was built around experimentation, comparison and controlled observation. The same approach will continue to guide future AKSIOM models.

If you experiment with the model, feedback, comparisons and unusual use cases are very welcome.

Have fun exploring it.