Krea 2 Turbo BBOX — GGUFs + W4A8 + INT8
Quantized by Rebel AI | @realrebelai
Take full control of your image compositions with Krea 2 Turbo BBOX, a specialized fine-tune of Krea 2 designed for layout-aware image generation.
This release brings the model to ComfyUI in multiple optimized formats, including GGUF quantizations, W4A8, and INT8 ConvRot, giving users more flexibility to balance image quality, memory consumption, and inference performance.
Unlike conventional text-to-image models, Krea 2 Turbo BBOX supports a structured prompting language that lets you guide where characters, objects, text, and comic panels appear in your generations.
Whether you're creating promotional flyers, gaming advertisements, cinematic scenes, comics, or complex multi-subject compositions, BBOX prompting provides greater control over image layout.
Available Model Formats
Q3_K_S GGUF High compression for memory-constrained GPUs
Q4_K_M GGUF Balanced compression and image quality
Q5_K_M GGUF Higher-fidelity quantization with moderate memory requirements
Q6_K GGUF High-quality GGUF with reduced quantization loss
Q8_0 GGUF Highest-precision GGUF option in this release
W4A8 4-bit weights with 8-bit activation quantization, designed for efficient inference with a compatible runtime
INT8 ConvRot 8-bit quantization using rotation-based techniques, targeting strong quality retention and efficient inference
All formats are based on the same Krea 2 Turbo BBOX fine-tune.
W4A8 and INT8 ConvRot are separate quantization approaches, not GGUF variants. Their actual performance depends on hardware, runtime, and kernel support.
Recommended Generation Settings
Sampling steps: 8
CFG: 1.0 (guidance disabled)
Architecture: Krea 2
Interface: ComfyUI
Prompting: Natural language or structured BBOX prompts
Sampler and scheduler configurations should follow a compatible Krea 2 Turbo workflow.
Installation — ComfyUI
GGUF Versions
Install or update ComfyUI-GGUF.
Download your preferred GGUF quantization.
Place the
.gguffile insideComfyUI/models/diffusion_models/.Restart ComfyUI.
Load the model using
UnetLoaderGGUFin a compatible Krea 2 Turbo workflow.
W4A8 and INT8 ConvRot Versions
Download the desired W4A8 or INT8 model.
Place the model in the directory expected by its compatible loader or workflow.
Install the appropriate quantized-model loader and required dependencies.
Select the matching model format and load it using the provided workflow.
Important: W4A8 and INT8 ConvRot require compatible inference support. They are not interchangeable with GGUF files, and the standard GGUF loader cannot load them directly.
Compatible workflows and format-specific requirements should be checked before use.
The diffusion model alone may not include the required text encoder, VAE, or other pipeline components.
How BBOX Prompting Works
Krea 2 Turbo BBOX uses a custom structured layout language that allows users to describe both image content and its intended spatial arrangement.
Approximately 90% of the fine-tune's training used structured layout prompts.
Instead of asking the model to place an astronaut somewhere near a rover, you can assign specific regions to the astronaut, rover, background objects, and text.
Bounding Box Coordinates
The model uses a 0–1000 coordinate system, independent of image resolution.
Format: [x0,y0,x1,y1]
x0, y0— Top-left cornerx1, y1— Bottom-right corner0,0— Top-left of the image1000,1000— Bottom-right of the image
Supported Prompting Tags
TagPurposeo[...]Place an object within a designated regiont[...]"TEXT"Place text in a specified areap[...]Define a comic or storyboard panelpe:1[...]Define a person with an identity IDac:1[...]Define an anime character with an identity IDfc:1[...]Define an anthropomorphic character with an identity ID
The numbered character tags can help describe recurring characters and identities.
Example BBOX Prompt — Astronaut
A photorealistic NASA-style space exploration photograph of an astronaut standing on an orange desert planet, holding a large black sign reading "Krea-2-Turbo BBOX". A giant ringed planet dominates the sky, an exploration rover rests nearby, and glowing blue crystals emerge from the ground.
@cinematic photography, realistic materials, dramatic sunlight, sharp focus; Photograph
~An orange rocky desert stretching toward distant mountains beneath an extraterrestrial sky.
pe:1[350,300,650,940] A lone astronaut wearing a detailed white spacesuit, standing upright and holding a rectangular black sign with both hands.
o[320,470,690,650] Large rectangular black sign held in front of the astronaut's chest.
t[350,510,660,610]"Krea-2-Turbo BBOX" bold white lettering centered on the black sign.
o[570,30,950,420] A gigantic ringed planet dominating the upper-right sky.
o[40,600,350,880] A futuristic exploration rover parked on the rocky desert.
o[720,580,950,960] Glowing blue crystalline formations emerging from the ground.These coordinates are learned spatial instructions rather than pixel-perfect placement guarantees.
For additional examples, see the developer's official prompting guide.
What Can You Create?
Promotional flyers: Control the placement of headlines, characters, and event details.
Gaming advertisements: Design esports graphics and promotional posters.
Comic books: Create multi-panel compositions and recurring characters.
Cinematic scenes: Arrange characters, objects, and backgrounds.
Product advertising: Position products and typography in designated regions.
Complex compositions: Organize multiple subjects with spatial instructions.
Choosing a Quantization
GGUF is a strong option for users looking to reduce model memory requirements and use ComfyUI-GGUF's supported quantized-model loading.
W4A8 offers more aggressive weight compression with 8-bit activations. It may provide efficient inference when supported by optimized kernels, although image quality can be affected by quantization.
INT8 ConvRot uses an 8-bit quantization approach with rotation-based transformations intended to improve quantization behavior. It may offer stronger quality retention than more aggressive 4-bit approaches, depending on the model and implementation.
Neither W4A8 nor INT8 is guaranteed to outperform GGUF in speed. GPU architecture, loader implementation, activation quantization, and kernel support all influence performance.
The goal of this release is to give users multiple options rather than force everyone into a single quantization format.
About This Release
These quantized versions were prepared by Rebel AI to make Krea 2 Turbo BBOX more accessible across different GPU configurations.
All formats preserve the original model's intended functionality while reducing numerical precision through different quantization methods.
Quantization does not add new training or change the model's underlying BBOX capabilities.
Image quality, text accuracy, and spatial adherence may vary between formats.
Credits
Original fine-tune: jimmycarter/krea2-turbo-bbox
Base model: Krea 2 by Krea
GGUF / W4A8 / INT8 quantizations: Rebel AI
Hugging Face: Krea 2 Turbo BBOX — Rebel AI
These are derivative quantized releases. Original model licensing and usage restrictions remain applicable. Review the upstream terms before commercial deployment.
Follow Rebel AI
YouTube: https://youtube.com/@realrebelai
X / Twitter: https://x.com/realrebelai
GGUF quantizations, W4A8 and INT8 experiments, low-VRAM ComfyUI workflows, benchmarking, and open-source AI research.
Rebel AI — Making powerful AI models accessible to everyone.
Description
gguf, int8, w4a8 formats available












