VisionForge

Inspiration

As a creator, I've always wanted to make visual novels and manga using AI. But every time I tried, I hit the same wall - my characters looked completely different in every image. Silver-haired Kaito in scene one became brown-haired in scene two. It was frustrating and unusable for any real storytelling. I built VisionForge to solve this problem once and for all.

What it does

VisionForge turns written stories into consistent AI artwork. You write your story with character descriptions, choose a visual style (Anime, Realistic, Sci-Fi, or Fantasy), and click "Forge Vision." The system extracts "Character DNA" - a structured JSON profile of each character's visual features - and uses it to generate perfectly consistent characters across all scenes. Export your story as Manga (black & white panels) or Manhwa (full color scroll).

How we built it

Used Gemini 2.0 Flash to parse user stories and identify characters
Generated character portraits using FIBO by Bria AI
Used Gemini Vision to extract detailed "Character DNA" as structured JSON
Locked the DNA and injected it into every scene generation for consistency
Built the web interface with Reflex (Python)
Added export features for Manga and Manhwa formats

Challenges we ran into

The biggest challenge was getting the DNA extraction right. I needed to capture enough visual detail to ensure consistency, but not so much that it became rigid or broke the generation. Finding the right balance took a lot of experimentation. Another challenge was managing API rate limits while keeping the user experience smooth with real-time progress indicators.

Accomplishments that we're proud of

I'm proud that VisionForge actually works! Characters genuinely look consistent across multiple scenes - something I couldn't achieve with any other tool. The "Character DNA" concept using FIBO's JSON-native architecture is elegant and effective. I'm also proud of the clean UI with the Electric Violet theme and the seamless export to Manga/Manhwa formats.

What we learned

I learned that structured data beats clever prompting. By extracting visual features as JSON and feeding them back systematically, I achieved consistency that prompt engineering alone couldn't deliver. FIBO's JSON-native design was perfect for this. I also learned how powerful agentic workflows can be - letting AI orchestrate the entire pipeline automatically.