Artificial Intelligence (AI) image generation technology is accelerating rapidly – in more ways than one. Recent developments have taken the industry from steady growth to relentless progress, now promising real-time, high-fidelity imaging.
It's not that these devices are slow – a minute is not too long to wait to “do more”. But users still want more: more realism, more versatility, more variety and more speed. And on the last point, researchers are happily presenting.
SDXL hits the gas pedal.
Stability AI has unveiled SDXL Turbo, which could represent a giant leap forward in AI image generation. We don't say this lightly: the recently announced model can generate images in 30 to 60 seconds longer than conventional generators take. It's almost, if not as effective, real-time AI image generation.
The SDXL Turbo is different from all previous Stable Diffusion models. Adversarial Diffusion Distillation (ADD) technology makes it possible to significantly reduce the number of steps required to generate high-quality images – even a single step can take as few as 30 to 100 steps for conventional images. “ADD is the first method to unlock single-step, real-time image fusion with base models,” Stability AI said in a research paper.
SDXL Turbo employs contrast training and output differentiation, streamlining the generation process and ensuring images are produced quickly while maintaining high fidelity.
As a result, the introduction of SDXL Turbo enables the immediate processing of complex and high-quality images. This new approach brings attention to the ganks, which have been largely forgotten since distribution technology began to dominate the scene.
Latent consistency models mean efficiency.
If you don't want to say goodbye to your “legacy” stable distribution models, however, researchers have a solution for you.
Latent coherence models (LCMs) and LCM-LoRA related to the development of the SDXL turbo each make unique contributions to the field.
LCMs, as described in their research paper, stand out for their ability to generate high-quality images by working effectively in the latent space of pre-trained autoencoders such as Stable Diffusion. LCM aims to increase the speed of image generation without significant loss in quality, focusing on high quality results. Using a one-step guided filtering method, LCM transforms pre-trained diffusion models into fast image generators, skipping unnecessary steps.
Practically speaking, users don't need to change anything else. Just download the model and use it as a standard SDXL scan. However, instead of going through so many steps, you can lower the parameter to the minimum. Instead of calculating the generation for each image for 25, 50 or 75 steps, the model produces good images in four steps in two seconds.
There are already great models with their own LCM versions for you to try. We recommend Hephaistos_NextGENXL for its versatility, but there are many great models to try out.
LCM-LoRAS: Starving any model
Released in conjunction with LCM, LCM-LoRA provides a universal velocity module that can be integrated into a variety of stable-diffusion models. “LCM-LoRA can be viewed as a plug-in neural PF-ODE solver with strong generalization capabilities,” the research paper reads.
LCM-LoRA is designed to enhance the efficiency of existing stable distribution models, making them faster and more versatile. It uses LoRA (low level adaptation) to update pre-trained weight matrices, reducing computational load and memory requirements.
With LCM-LoRA, conventional steady-state models experience a significant increase in their image generation speed, making them highly effective for a variety of tasks. Users don't even need to download a new model – just enable LCM LoRA and create images as fast as LCM Mode.
LCM-LoRAs can be downloaded here for SD 1.5 and SDXL.
Quality and speed
Despite these technological leaps, the need to balance speed and image quality remains. Fast-generation tools like SDXL Turbo and LCM-LoRA speed up the creative process, but they do so at the expense of some image fidelity. In other words, an image generated with 50 layers and a good model will always have higher quality or image fidelity than an image generated with 5 layers and a good LCM model.
However, this trade-off is reduced in terms of serviceability in typical workflows where multiple images are created to find the perfect one. Subsequent iterations such as image-to-image or colorization can improve detail on these first-cut images, compensating for the initial loss of quality. A properly calibrated image generated by one of these fast technologies can be as good as an image generated by a conventional Stable Diffusion model.
Fasten your seat belts because the AI image generation space is turning into overdrive – and few people need speed more than AI enthusiasts.
Stay on top of crypto news, get daily updates in your inbox.
