Average Ratings 0 Ratings
Average Ratings 0 Ratings
Description
The Bonsai Image Ternary 4B MLX 2-bit is a text-to-image diffusion transformer specifically designed for deployment on Apple Silicon, emphasizing quality in its Bonsai Image variant. This model utilizes ternary weights of {−1, 0, +1} along with FP16 group-wise scaling in its transformer layers, which encompass Q/K/V projections, output projections, and MLP weights. Notably, it reduces the size of the FLUX.2 Klein 4B transformer from 7.75 GB FP16 to just 1.21 GB, achieving a remarkable 6.4× smaller footprint while maintaining visual quality and fidelity to prompts akin to the original model. The deployment package for Apple Silicon is 3.88 GB, which includes the MLX 2-bit diffusion transformer, a 4-bit Qwen3-4B text encoder, and an FP16 Flux2 VAE. After the text encoder handles prompt encoding, it is offloaded to ensure that only the compact transformer and VAE remain in memory during the denoising loop. Furthermore, the model employs a 4-step FlowMatchEuler sampler with guidance set at 1.0 and a shift of 3.0, eliminating the need for CFG and negative prompts, thus streamlining the generation process for enhanced user experience. Overall, this innovation represents a significant advancement in efficient and effective image generation technology.
Description
Karlo serves as an innovative model designed to create images from textual descriptions. It enhances the impressive unCLIP architecture developed by OpenAI by improving the conventional super-resolution model, enabling it to capture complex details at an impressive resolution of 256px, while effectively reducing noise through a limited number of denoising iterations.
In developing Karlo, we undertook a comprehensive training regimen that began from the ground up, leveraging a substantial dataset of 115 million image-text pairs, which included COYO-100M, CC3M, and CC12M. For the Prior and Decoder sections, we utilized the advanced ViT-L/14 text encoder sourced from OpenAI's CLIP library. To boost performance, we implemented a notable alteration to the original unCLIP design; rather than using a trainable transformer in the decoder, we opted to incorporate the text encoder from ViT-L/14, thereby enhancing the model's capability. This strategic choice not only streamlined the architecture but also contributed to improved image quality and fidelity.
API Access
Has API
Yes
API Access
Has API
No
Integrations
B^ DISCOVER
No
B^ EDIT
No
Pricing Details
No price information available.
Free Trial
No
Free Version
No
Pricing Details
Free
Open source
Free Trial
No
Free Version
Yes
Deployment
Web-Based
No
On-Premises
Yes
iPhone App
Yes
iPad App
Yes
Android App
Yes
Windows
Yes
Mac
Yes
Linux
Yes
Chromebook
No
Deployment
Web-Based
Yes
On-Premises
Yes
iPhone App
No
iPad App
No
Android App
No
Windows
No
Mac
No
Linux
No
Chromebook
No
Customer Support
Business Hours
No
Live Rep (24/7)
No
Online Support
Yes
Customer Support
Business Hours
No
Live Rep (24/7)
No
Online Support
No
Types of Training
Training Docs
Yes
Webinars
No
Live Training (Online)
Yes
In Person
No
Types of Training
Training Docs
Yes
Webinars
No
Live Training (Online)
No
In Person
No
Vendor Details
Company Name
PrismML
Founded
2026
Country
United States
Website
prismml.com
Vendor Details
Company Name
Kakao Brain
Founded
2017
Country
South Korea
Website
github.com/kakaobrain/karlo