I trained two custom LoRAs on 73 of my own ink drawings and made a short film with them — full process included

An independent artist has demonstrated a sophisticated, end-to-end pipeline for creating AI-generated short films using custom-trained models. For the Arca Gidan Prize—an open-source AI film contest—they produced 'INNECENCE' by training two distinct LoRAs (Low-Rank Adaptations) on a personal dataset of 73 hand-drawn Chinese ink illustrations. The first, a Z-Image LoRA (rank 32), was trained for 80 epochs to capture the core artistic style. The second, an LTX-V 2.3 LoRA (rank 64), was trained for motion generation. Both were trained using Musubi-tuner on a RunPod H100 cloud instance, with the artist noting that practical sample quality was more important than perfect loss curves.

The production workflow was multi-stage and highly technical. It began with generating keyframes using the custom Z-Image LoRA. These frames were then refined for art direction using QwenImageEdit. The LTX-V 2.3 model, fine-tuned with its own LoRA, was used for image-to-video generation, creating both animated shots and specific ink-wash transition effects—a process that required two generation passes per shot. The final footage was upscaled to HD using SeedVR2.5 and assembled in the video editor Kdenlive. The artist has openly shared all components of their pipeline, including training guides, captioning scripts, and ComfyUI workflows, providing a valuable blueprint for other creators looking to personalize AI video generation.