Announcement: NovelAI Diffusion V5 Full now has a "Medium" and "High" Effort Toggle

Today we’re introducing effort levels for NovelAI Diffusion V5 Full. A new toggle lets you choose between “Medium” effort, a new, more cost-efficient option, and “High” effort, which is the current behaviour at the same cost to the Opus Tier usage limits and paid Anlas as before.

Compared to a “high” effort 23 step generation, a “medium” effort generation is more efficient and costs about 42% less usage at default settings, with quality that held up against the original in our internal blind tests.

To build the new “medium” effort level, we have distilled the original model to make it run at only 14 steps, while delivering high image quality and preserving the original model’s ability to accurately follow prompts.

In our internal blind tests, “high” and “medium” effort levels have both been found to deliver generations of roughly equal quality.

However, due to how the distillation was performed, “medium” effort does not support changing certain settings:

The step count is locked to 14 steps, exactly half of the maximum step count available within the Opus subscription limits. Additionally, the sampler cannot be changed from Euler Ancestral, and there is no way to specify a custom Undesired Content prompt, or change its preset. While Prompt Guidance can still be adjusted as normal, Prompt Guidance Rescale is not supported.

If you wish to do negative prompting on the new “medium” effort level, please try using negative emphasis. For example, if a character is usually wearing a hat and you would like to remove it, please try adding “-3::hat::” to your prompt, no Undesired Content needed.

Here’s a gallery of examples comparing the performance between Medium & High Effort!

The new “medium” effort level is available for both NovelAI Diffusion V5 Full and its inpainting mode.