The Elusive Digital Eraser: AI Struggles to Remove Hands, Preserve Custom Mascot in Video Editing

By Investigative Tech Journalist
An analysis of emerging AI video editing challenges

In the rapidly evolving field of AI-powered content creation, a new frontier has emerged: the seamless removal of objects from video. However, a detailed case study from a digital creator reveals that this promise is fraught with technical complexity, especially when the task involves preserving a unique, custom-designed object. The creator's goal was deceptively simple—edit out their own hands from a video where they were moving a mascot, making it appear as if the mascot was moving autonomously. The reality of achieving this with current AI tools proved to be a significant obstacle.

The creator, posting under the username degel12345 on a popular Stable Diffusion forum, employed a sophisticated pipeline. The workflow involved using the Wan 2.1 VACE 14B model for inpainting—the process of regenerating masked areas of an image or video—combined with a specialized Low-Rank Adaptation (LoRA) model. This LoRA was painstakingly trained on approximately 18 images of the unique mascot, dubbed "florbus," to teach the AI its specific appearance. The video was segmented, hands were masked using the SAM3 segmentation tool, and a clean background plate was used as a reference.

The Core of the Problem: AI's Lack of Context

Despite this meticulous setup, the results were deeply flawed. The AI failed to coherently reconstruct the mascot where the creator's hands had been. Instead, it produced bizarre artifacts: a thumb appeared to create a hole in the mascot's body, other fingers were merely covered with mascot-like texture, and extraneous, non-existent dolphin body parts were generated. The mascot's form was not preserved, demonstrating the AI's fundamental lack of understanding of the object it was supposed to be reconstructing.

This scenario exemplifies the classic definition of a technical problem. According to standard references, a problem is "something that causes difficulty or that is hard to deal with." Merriam-Webster similarly defines it as "a question raised for inquiry, consideration, or solution" or "an intricate unsettled question." The creator's experience is a textbook case of an intricate, unsettled question at the cutting edge of generative AI.

"The problematic part is the mascot shape - the model have no idea how the mascot looks like so it can't be consistent across generations," the creator wrote, pinpointing the AI's core limitation. The LoRA, intended to solve this, showed weak influence, requiring very specific prompts like "florbus dolphin toy" to work at all, rather than just the trigger word "florbus."

Technical Tightrope: Balancing Removal and Reconstruction

The investigation into the creator's settings reveals a delicate balancing act. They used a VACE (Video Attention Control and Editing) strength of 1.5 to ensure the hands were removed. When this strength was lowered to 1.0, the hands remained fully visible. However, at 1.5, the aggressive removal process destroyed the contextual information needed to accurately rebuild the mascot behind the hands. The AI, lacking a robust internal model of "florbus," defaulted to generating generic or distorted shapes based on its broader training data, which includes concepts like "dolphin toy."

This highlights a critical gap in current AI video editing suites. While tools for segmentation and background replacement are becoming robust, the intelligent, context-aware reconstruction of occluded unique objects remains a major hurdle. The AI can remove content, but it cannot reliably infer what *should* be there if that thing is not already a well-defined concept within its vast but generalized dataset.

Broader Implications for AI Content Creation

This case is not an isolated incident but rather a symptom of a broader challenge. As creators and businesses seek to use AI for professional video post-production—removing boom mics, erasing modern objects from period pieces, or cleaning up product shots—the need for AI to understand and preserve specific, custom elements is paramount. The current generation of models excels at generalization but falters at precise, customized reconstruction.

The creator's arduous process of training a LoRA represents the community's current best workaround: attempting to inject specialized knowledge into the model. Yet, as seen here, integrating this specialized knowledge seamlessly into a complex video inpainting workflow is non-trivial. The interaction between the inpainting model's strength, the LoRA's weight, the mask precision, and the prompt phrasing creates a high-dimensional parameter space that is difficult to navigate to a perfect solution.

The Path Forward

Solving this class of problems will likely require advancements in several areas. First, more robust and easier-to-train methods for teaching AI about custom objects are needed. Second, inpainting models need better mechanisms for incorporating reference imagery and maintaining spatial consistency across video frames. Third, user interfaces must evolve to give creators more intuitive control over the reconstruction priority—specifying what to remove versus what to protect and rebuild.

For now, creators like degel12345 are left on the frontier, experimenting with sampler settings, LoRA strengths, and prompt engineering. Their work, while facing a clear and difficult problem as defined by any standard dictionary, provides invaluable real-world stress tests for these technologies. Each failed generation and forum post detailing the issue contributes to the collective understanding that will eventually lead to more reliable and powerful AI editing tools. The digital eraser is not yet perfect, but the struggle to refine it is actively defining the next capabilities of creative AI.

This report is based on an analysis of technical discussions within the AI development community and standard definitions of technical obstacles.