https://liu-yufei.github.io/tags/object-removal/Object RemovalHugo Blox Builder (https://hugoblox.com)en-usThu, 01 Jan 2026 00:00:00 +0000https://liu-yufei.github.io/media/icon_hu17033853472409137288.pnghttps://liu-yufei.github.io/tags/object-removal/https://liu-yufei.github.io/project/wan-vace-object-removal/Thu, 01 Jan 2026 00:00:00 +0000https://liu-yufei.github.io/project/wan-vace-object-removal/<p>This Huawei collaboration project addresses the shortage of high-quality paired data for video object removal, especially when shadows and reflections must also be removed.</p> <p>I independently built a data generation and fine-tuning pipeline using Wan2.2-5B, SAM2, and a baseline removal model. From roughly 4,000 generated groups, I selected 1,070 high-quality training pairs and retained 2,500 difficult cases for evaluation. I then fine-tuned the VACE module of Wan2.1-VACE-1.3B with DiffSynth-Studio, Accelerate, and DeepSpeed on six NVIDIA A800 GPUs.</p> <style> .wan-comparisons { margin: 3.25rem auto; max-width: 1120px; } .wan-comparisons h2 { margin: 0 0 1.5rem; font-size: 1.75rem; } .wan-comparison { --split: 50%; position: relative; width: 100%; aspect-ratio: 40 / 23; overflow: hidden; border-radius: 0.8rem; background: #111; } .wan-comparison + h2 { margin-top: 3.5rem; } .wan-comparison video { position: absolute; inset: 0; width: 100%; height: 100%; object-fit: contain; } .wan-compare-after { clip-path: inset(0 0 0 var(--split)); } .wan-comparison-label { position: absolute; top: 1rem; z-index: 4; padding: 0.35rem 0.7rem; border-radius: 999px; color: #fff; background: rgba(0, 0, 0, 0.62); font-size: 0.85rem; font-weight: 700; backdrop-filter: blur(6px); } .wan-comparison-label.left { left: 1rem; } .wan-comparison-label.right { right: 1rem; } .wan-comparison-line { position: absolute; z-index: 3; top: 0; bottom: 0; left: var(--split); width: 2px; background: #fff; transform: translateX(-1px); pointer-events: none; } .wan-comparison-handle { position: absolute; z-index: 4; top: 50%; left: var(--split); width: 2.65rem; height: 2.65rem; border: 2px solid #fff; border-radius: 50%; color: #fff; background: rgba(0, 0, 0, 0.55); transform: translate(-50%, -50%); display: grid; place-items: center; font-size: 1rem; pointer-events: none; } .wan-comparison-range { position: absolute; z-index: 5; inset: 0; width: 100%; height: 100%; margin: 0; opacity: 0; cursor: ew-resize; } .wan-comparison-note { margin: 0.75rem 0 0; text-align: center; font-size: 0.88rem; opacity: 0.68; } @media (max-width: 700px) { .wan-comparison-label { top: 0.6rem; padding: 0.25rem 0.55rem; font-size: 0.72rem; } .wan-comparison-label.left { left: 0.6rem; } .wan-comparison-label.right { right: 0.6rem; } .wan-comparison-handle { width: 2.2rem; height: 2.2rem; } } </style> <section class="wan-comparisons" aria-labelledby="wan-comparison-title"> <p class="wan-comparison-note">All videos are displayed at 480 × 276, matching the output resolution of our model.</p> <h2 id="wan-comparison-title">Condition Video vs. Baseline</h2> <div class="wan-comparison" data-video-compare data-video-sync data-sync-duration="3.0625"> <video data-sync-master muted playsinline preload="auto" poster="/project/wan-vace-object-removal/condition-poster.png?v=resolution-matched-20261007"><source src="https://liu-yufei.github.io/project/wan-vace-object-removal/condition.mp4?v=resolution-matched-20261007" type="video/mp4"></video> <video class="wan-compare-after" muted playsinline preload="auto" poster="/project/wan-vace-object-removal/baseline-poster.png?v=resolution-matched-20261007"><source src="https://liu-yufei.github.io/project/wan-vace-object-removal/baseline.mp4?v=resolution-matched-20261007" type="video/mp4"></video> <span class="wan-comparison-label left">Condition Video</span> <span class="wan-comparison-label right">Baseline</span> <span class="wan-comparison-line"></span><span class="wan-comparison-handle">↔</span> <input class="wan-comparison-range" type="range" min="0" max="100" value="50" aria-label="Compare the condition video with the baseline result"> </div> <p class="wan-comparison-note">Drag the slider to compare both videos frame by frame.</p> <h2>Condition Video vs. Ours</h2> <div class="wan-comparison" data-video-compare data-video-sync data-sync-duration="3.0625"> <video data-sync-master muted playsinline preload="auto" poster="/project/wan-vace-object-removal/condition-poster.png?v=resolution-matched-20261007"><source src="https://liu-yufei.github.io/project/wan-vace-object-removal/condition.mp4?v=resolution-matched-20261007" type="video/mp4"></video> <video class="wan-compare-after" muted playsinline preload="auto" poster="/project/wan-vace-object-removal/ours-poster.png?v=resolution-matched-20261007"><source src="https://liu-yufei.github.io/project/wan-vace-object-removal/ours.mp4?v=resolution-matched-20261007" type="video/mp4"></video> <span class="wan-comparison-label left">Condition Video</span> <span class="wan-comparison-label right">Ours</span> <span class="wan-comparison-line"></span><span class="wan-comparison-handle">↔</span> <input class="wan-comparison-range" type="range" min="0" max="100" value="50" aria-label="Compare the condition video with our result"> </div> <p class="wan-comparison-note">Drag the slider to inspect object and shadow removal.</p> </section> <p>Human evaluation found 2,446 of the 2,500 difficult cases usable, a 97.84% success rate, with improved object-shadow removal and background reconstruction.</p>