Stable Diffusion maker Stability AI has unveiled Stable Doodle, a sketch-to-image tool to convert simple drawings into high-quality images.

Developed by AI-based image editing platform Clipdrop and Stability AI, Stable Doodle can be accessed for free on the Clipdrop by Stability AI website, along with the latest Stable diffusion model SDXL 0.9. In March, Stability AI acquired Init ML, the creator of Clipdrop.

Stable Doodle is designed to cater to both experienced users and beginners, regardless of their familiarity with AI tools. By harnessing the power of Stable Doodle, anyone with basic drawing skills and internet access can generate high-quality original images within seconds.

Stable Doodle  allows for artistic customisation, offering 14 styles to choose from via Stable Diffusion XL. These styles range from realistic photography to cinematic aesthetics to imaginative fantasy art and origami-inspired designs.

This is not the first time that we have a sketchy cousin of Stable Diffusion. Earlier, an engineer from Replicate, who goes by the GitHub name zeke, developed Scribble Diffusion to convert hand-drawn artwork, along with an accompanying text prompt, into a new art. 

Decoding the Engineering of Stable Doodle

Stable Doodle combines the image-generation technology of Stability AI’s Stable Diffusion XL with the formidable T2I-Adapter. Developed by Tencent ARC (license), the T2I-Adapter is a precise condition control solution that enhances AI image generation.

By introducing trainable parameters to existing large diffusion models, the T2I-Adapter allows for the incorporation of additional input conditions like sketches, segmentation maps, or key poses. 

This framework supports multiple models for input guidance simultaneously, granting enhanced control over the generation process. In the context of Stable Doodle, the T2I-Adapter supplements the pre-trained text-to-image model (SDXL), enabling it to comprehend sketch outlines and produce images based on prompts combined with the defined outlines.

The T2I-Adapter network consists of approximately 77 million parameters, delivering additional guidance to pre-trained text-to-image (SDXL) models while maintaining the integrity of the original large text-to-image models.

Mostaque is Always on the Go

At the Bloomberg Technology Summit held in San Francisco, Stability AI’s CEO, Emad Mostaque, acknowledged the concerns surrounding the creation of realistic AI-generated deepfakes during an on-stage interview. Mostaque disclosed that the company had developed “photo-realistic models” but decided against releasing them at that time due to various considerations. He stressed the importance of implementing features such as watermarking to establish standards that enable tracking and appropriate usage of AI-generated content.

Recently, Mostaque gained attention again for his statement during an interview with Peter H. Diamandis for the Moonshots and Mindsets Podcast. He claimed that within the next five years, human programmers would become obsolete and that 41% of code on platforms like GitHub is generated by AI. However, some users have pointed out that there is no data available to support this assertion.