Many powerful image models can generate stunning visuals, but they are often difficult to customize for specific creative needs. Released by KreaAI, Krea 2 is a 12-billion-parameter open source text-to-image diffusion model designed to generate images directly from natural-language descriptions.
The model is built for a wide range of applications, including-Creative exploration,Commercial design workflows , Visual production, Developer integrations ,Research experiments. Instead of being only a generation model, Krea 2 is designed as a flexible foundation that can fit into different creative ecosystems. The detailed insights can be found be accessing their research paper.
Basically Krea team released two variants-
(a) Krea2 Raw (Base)-The Raw checkpoint is designed primarily as a foundation model rather than a final image-generation model. It is not recommended for direct inference/image gen use because its main purpose is customization and further training. Train your loras on the top of it and use with Turbo variant. This makes it useful for- Fine-tuning, Domain-specific training, Creating custom LoRAs, Research experiments.
 |
| Krea2 Raw/Base Showcase |
(b) Krea 2 Turbo- It is an 8-step distilled checkpoint built for fast, high-quality text-to image generation. The model is distilled from a fully post trained Reinforcement Learning (RL) checkpoint, allowing it to produce better-quality images compared to Raw while maintaining much faster generation. The trade-off is that Turbo produces less variation between outputs, making it more consistent but slightly less flexible.
 |
| Krea2 Turbo Showcase |
Installation
1. First, install ComfyUI if you are a new user. Older user required to do the ComfyUI update from the Manager.
2. All Krea2 base and Turbo (Bf16/fp8/Nvfp4/Int8) model listed below from community. Choose any of the model for you use case and system resources. Save any of them into ComfyUI/models/diffusion_models folder.
These are Krea2 Model repacked officially by ComfyUI Team:
| Sl No. | Model | Format | Description (VRAM Support) |
|---|---|---|---|
| 1 | krea2_raw_bf16.safetensors | BF16 | Full precision Krea 2 Raw model. Provides maximum image quality and accuracy. Requires higher VRAM, suitable for GPUs with 16GB+ VRAM. |
| 2 | krea2_raw_fp8_scaled.safetensors | FP8 Scaled | Quantized Krea 2 Raw version with reduced memory usage while maintaining good quality. Suitable for GPUs with around 8GB–12GB VRAM. |
| 3 | krea2_turbo_bf16.safetensors | BF16 | Turbo variant optimized for faster generation with BF16 precision. Requires higher VRAM, recommended for GPUs with 16GB+ VRAM. |
| 4 | krea2_turbo_fp8_scaled.safetensors | FP8 Scaled | Memory-efficient Turbo model using FP8 scaling. Designed for faster inference on mid-range GPUs with approximately 8GB–12GB VRAM. |
| 5 | krea2_turbo_mxfp8.safetensors | MXFP8 | Mixed FP8 quantized Turbo model optimized for efficiency. Suitable for GPUs with lower VRAM requirements, generally around 8GB+ VRAM. |
| 6 | krea2_turbo_nvfp4.safetensors | NVFP4 | Highly compressed Turbo model using NVIDIA FP4 format. Designed for low VRAM usage, suitable for GPUs with approximately 6GB–8GB VRAM. |
These are Krea2 variants quantized be developer Winnougan:
| Sl No. | Model | Format | Description (VRAM Support) |
|---|---|---|---|
| 1 | krea2_base_fp8.safetensors | FP8 Tensorwise | Recommended for NVIDIA RTX 40xx series GPUs. |
| 2 | krea2_base_mxfp8.safetensors | MXFP8 | Recommended for NVIDIA RTX 50xx Blackwell GPUs. |
| 3 | krea2_base_nvfp4.safetensors | NVFP4 | Optimized for RTX 50xx Blackwell GPUs only. |
| 4 | krea2_base_int8.safetensors | INT8 Row-wise | Recommended for NVIDIA RTX 30xx series GPUs. |
| 5 | krea2_base_convrot_int8.safetensors | INT8 + ConvRot | Best quality INT8 version, recommended for RTX 30xx GPUs. |
| 6 | krea2_turbo_fp8.safetensors | FP8 Tensorwise | Recommended for NVIDIA RTX 40xx series GPUs. |
| 7 | krea2_turbo_mxfp8.safetensors | MXFP8 | Recommended for NVIDIA RTX 50xx Blackwell GPUs. |
| 8 | krea2_turbo_nvfp4.safetensors | NVFP4 | Optimized for RTX 50xx Blackwell GPUs only. |
| 9 | krea2_turbo_int8.safetensors | INT8 Row-wise | Recommended for NVIDIA RTX 30xx series GPUs. |
| 10 | krea2_turbo_convrot_int8.safetensors | INT8 + ConvRot | Best quality INT8 version, recommended for RTX 30xx GPUs. |
3. Download text encoder (qwen3vl_4b_bf16.safetensors / qwen3vl_4b_fp8_scaled.safetensors). Choose any of them. save this into ComfyUI/models/text_encoders folder.
4. Download Vae (qwen_image_vae.safetensors) and save this into ComfyUI/models/vae folder.
5. Refresh and restart comfyui to take effect.
Workflow
1. Download Krea 2 Base/Turbo workflow (Krea2-basic.json) from our hugging face repository. Use same workflow for Krea2 Base/Turbo model.
2. Drag and drop the workflow into ComfyUI.
3. Load the models(Krea2 Base/Turbo, text encoders, vae) into its relevant nodes.
4. Set KSampler configuration -
Krea2 Raw/Base:
Resolution- up to 1024
Steps-52
CFG- 3.5
Sampler: er_sde
Scheduler- simple
Krea2 Turbo:
Resolution-1024 to 2048
Steps-8
CFG- 0-1(dsiabled)
Sampler: er_sde
Scheduler- simple
5. Put positive prompts into prompt box. use natural prompting with detailed style to get the best out of it.
6. Hit run to start image generation.
Krea 2 represents a useful shift in open source image generation. Instead of treating customization and usability as competing goals, it separates them into two connected stages. The Raw model gives creators room to experiment and build specialized capabilities, while Turbo focuses on delivering fast and polished results.
This approach feels closer to how real creative workflows operate experiment first, then move into efficient production. As more open-source image models adopt similar strategies, the future may not just be about having bigger models, but about creating better ecosystems around them.
 High Quality Image Gen){kind=link}