Custom LoRAs in 2026: How 73 Ink Drawings Powered an AI Short Film with Stable Diffusion

Custom LoRAs in 2026: How 73 Ink Drawings Powered an AI Short Film with Stable Diffusion

Custom LoRAs trained on 73 hand-drawn Chinese ink illustrations have powered the creation of a compelling AI-generated short film titled Innocence, showcasing the potential of personalized machine learning in digital art. The project, submitted to the open-source Arca Gidan Prize under the theme of "Time," represents a novel fusion of analog craftsmanship and generative AI, proving that minimal, static datasets can yield rich, motion-driven narratives when paired with precise training techniques.

Training Custom LoRAs on Ink Drawings

The filmmaker, known online as xCaYuSx, trained two specialized LoRAs—Z-Image and LTX-V 2.3—using Musubi-tuner on a RunPod H100. Each model was fine-tuned on the same 73-image dataset of 20-year-old ink drawings, avoiding video or motion data entirely. Despite noisy loss curves and no clear plateau during training, inference results revealed a distinct, recognizable aesthetic: delicate ink washes, fluid brushstrokes, and atmospheric textures that mirrored the artist’s original style.

This approach defies conventional wisdom: Stable Diffusion fine-tuning doesn’t require thousands of samples. With careful captioning (generated via Qwen3-VL), even 73 images can capture nuanced artistic intent. The Z-Image LoRA learned to preserve ink bleed and paper grain, while LTX-V 2.3 mastered temporal motion patterns.

Animating with LTX-V 2.3 and Z-Image LoRA

The animation pipeline was meticulously engineered. Z-Image generated static keyframes, which were then refined using QwenImageEdit for compositional control. LTX-V 2.3 handled image-to-video transitions, producing ink-wash effects that mimicked the slow drying and bleeding of traditional ink on rice paper.

Each shot required two generations: one for the animated still and another for the transition, with SeedVR2.5 applied for high-definition upscaling. AI frame interpolation was critical—without it, motion felt robotic. Prompting for an "ink-wash reveal" often resulted in artifacts like visible paintbrushes or generic crossfades. Through iterative seed variation and prompt engineering, the artist achieved subtle, painterly motion that felt organic.

The Arca Gidan Prize Submission Process

Final assembly was completed in Kdenlive, ensuring seamless pacing and emotional rhythm. With over 90 entries in the Arca Gidan competition, Innocence stands out for its emotional depth and stylistic coherence, achieved without commercial models or proprietary tools.

All training guides, ComfyUI workflows, and captioning scripts have been openly shared on the Arca Gidan submission page, making this a rare case of end-to-end transparency in AI filmmaking. The filmmaker emphasizes that quality should be judged by visual output, not loss metrics alone—a lesson for practitioners relying solely on numerical convergence.

Before and After: The Ink Wash Transformation

Before: Static ink drawings on rice paper, no motion, limited to gallery display.

After: Dynamic sequences where ink bleeds across frames, brushstrokes breathe, and shadows shift like wind through paper—all generated from a 73-image dataset using custom LoRAs and Stable Diffusion.

Why This Matters for Open-Source Art

Projects like Innocence underscore a broader trend: the democratization of cinematic expression through personalized AI models trained on intimate, human-made datasets. As AI tools evolve, artists aren’t just prompting machines—they’re teaching them their language. This isn’t just generative animation; it’s algorithmic apprenticeship.

Custom LoRAs trained on hand-drawn ink drawings have not only produced a visually arresting short film but have redefined what’s possible when personal artistry meets algorithmic precision.