NovelAI Diffusion V5 is here, and it’s a different class of model.

Sharper natural language understanding. Expanded multi-language prompting. Reworked character positioning. Higher image fidelity. This is the most advanced model we’ve ever shipped.

Both V5 Curated and V5 Full are available today

What Makes V5 Different?

V5 isn’t a big deal because of any single feature. It’s a big deal because of how all of them work together. Better prompt understanding means better compositions. A better VAE means cleaner details and cleaner alpha transparency. And better character positioning means comic generation actually works the way you’d want it to. Every upgrade compounds the others; this is the most refined, most capable model we’ve built.

Here’s what’s new:

More Characters & Improved Character Positioning

22 characters. One scene.
V5 renders up to 22 distinct characters in a single scene, with positioning to match. Place characters anywhere on your canvas and V5 follows, giving you greater control over composition.

Generate Full Comic Pages

Full comic pages, zero stitching.
Describe your panel layout, characters, and composition in natural language, and V5 generates the entire page at once. No more generating panels separately and stitching them together yourself.

Image Quality

More channels means more details.
Thanks to the larger model size, improved model architecture, and the new 32-channel VAE, smaller details like eyes, jewelry, text, and accessories are rendered with meaningfully higher accuracy.

Backgrounds Improved + Transparency

Better details didn’t stop at the character level.
With V5, environments, architecture, and foliage all render with real detail instead of fading into the backdrop. Plus, adding transparent backgrounds is now effortless.

Multi-Language Expansion & Natural Language Improvements

Describe it your way, in your language.
English and Japanese are fully supported, and during testing we saw solid results in Chinese, German, Spanish, and Portuguese as well. Write your prompts in natural sentences (or tags still work if that’s more your style)

Text Rendering, Reworked

Now supports up to 750 characters of text.
V5 renders text in multiple languages, and it’s easier to trigger than ever. Just quote your text and the frontend automatically builds the “Text:” block for you (previously a manual step in V4.5). Prefer to hand-build it yourself? That’s still fully supported.

Frequently Asked Questions

Is V5 replacing V4.5?
No, V4.5 is still available to use.

Do I need to change how I prompt to use V5?
No. Tag-based prompting still works exactly as it did. Natural language prompting just got a lot more capable if you want to use it.

What happened to Vibe Transfer, and Precise Reference?
They’re not part of the current launch but are planned as post-launch additions for V5, consistent with how we’ve rolled out past model generation additions.

Which languages can I prompt in?
English and Japanese are officially supported. Chinese, German, Spanish, and Portuguese have shown usable results in testing but weren’t a primary training focus, so results may be less consistent.

What subscription tiers or users are eligible to use the new V5 model?
V5 is available to all users and subscription tiers.