Research map / Video / Video generation
MiniMax-Remover: Taming Bad Noise Helps Video Object Removal
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MiniMax-Remover, a two-stage DiT-based video object removal model that drops text cross-attention for learnable contrastive condition tokens, then distills it on 10K curated videos with a minimax "bad noise" adversarial training scheme, achieving strong removal with 6 steps and no CFG.
Problem
Video object removal suffers from hallucinated objects and artifacts, and existing diffusion methods are slow because they need many sampling steps and classifier-free guidance.
Approach
Remove text cross-attention from Wan2.1-1.3B and inject learnable positive/negative condition tokens via self-attention; then apply minimax training where adversarial "bad" noise (sign-gradient perturbation of noise) is searched and the model is trained to remove objects from it, enabling CFG-free 6-step inference.
Key results
- On DAVIS, Ours (50 steps) reaches SSIM 0.9847 and PSNR 36.66 versus the best baseline DiffuEraser at 0.9818 and 34.42; the 6-step variant reaches 0.9842 and 36.56.
- GPT-o3 removal success rate on DAVIS is 91.11% for the 50-step model (82.22% at 6 steps) versus 56.67% for the best prior method (ProPainter and DiffuEraser).
- Inference takes about 0.18s per video in the efficiency table (1.05B params, 6 steps, no CFG, 8.2GB GPU memory), versus 0.35s for DiffuEraser at 70 steps with CFG.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace prompt conditioning with learnable contrastive tokens (ablation) | Stage 1 video object remover · DAVIS | GPT-O3 success rate: 51.11 -> 56.67 | Helped |
| remove cross-attention while retaining token conditioning (ablation) | Stage 1 video object remover · DAVIS | GPT-O3 success rate: 56.67 -> 53.33 | Hurt |
| inject contrastive tokens into self-attention after removing cross-attention (ablation) | Stage 1 video object remover · DAVIS | GPT-O3 success rate: 53.33 -> 71.11 | Helped |
| use human-curated data instead of WebVid data with random noise (ablation) | Stage 2 video object remover · DAVIS | GPT-O3 success rate: 65.56 -> 72.22 | Helped |
| add adversarial bad noise to human-curated training data (ablation) | Stage 2 video object remover · DAVIS | GPT-O3 success rate: 72.22 -> 91.11 | Helped |
| remove CFG at inference (ablation) | MiniMax-Remover · DAVIS | GPT-O3 success rate: 90.00 -> 91.11 | Helped |
| remove CFG at inference (ablation) | MiniMax-Remover · DAVIS | SSIM: 0.9853 -> 0.9847 | Hurt |
| replace random noise with adversarial-based noise (ablation) | Stage 2 video object remover · DAVIS | GPT-O3 success rate: 72.22 -> 91.11 | Helped |
| replace inversion-based noise with adversarial-based noise (ablation) | Stage 2 video object remover · DAVIS | GPT-O3 success rate: 70.00 -> 91.11 | Helped |
| use 6 sampling steps instead of 50 | MiniMax-Remover · DAVIS | GPT-O3 success rate: 91.11 -> 82.22 | Hurt |
| use 6 sampling steps instead of 50 | MiniMax-Remover · Pexels | GPT-O3 success rate: 81.0 -> 76.5 | Hurt |
| use 6-step MiniMax-Remover instead of the strongest prior baseline | video object removal methods · DAVIS | GPT-O3 success rate: 56.67 -> 82.22 | Helped |
Surprising findings
- Replacing text prompts with learnable contrastive tokens showed no significant performance drop, suggesting the text encoder is redundant for object removal.
- Inversion-based noise did not outperform random noise on success rate, and the paper finds it non-Gaussian.
Limitations
- Stage 2 remover trained on only 10K videos, limiting robustness and generalization.
- VAE encoding and decoding account for a large portion of computation.
- Stage 1 without minimax still has artifacts and occasional undesired object regeneration.
Experimental setup
Datasets: WebVid-10M Pexels (17K videos, 10K curated) DAVIS
Benchmarks: DAVIS Pexels (200 videos)
Metrics: SSIM PSNR temporal consistency (CLIP-based) GPT-o3 visual quality score GPT-o3 success rate user preference
Baselines: ProPainter VideoComposer COCOCO FloED DiffuEraser VideoPainter VACE
Builds on: Wan2.1 classifier-free guidance rectified flow Senorita-Remover mask strategy Grounded-SAM2
Closest papers by meaning
- Classic Video Denoising in a Machine Learning World: Robust, Fast, and Controllable (CVPR 2025 · similarity 0.86)
- RORem: Training a Robust Object Remover with Human-in-the-Loop (CVPR 2025 · similarity 0.86)
- FreeLong: Training-Free Long Video Generation with SpectralBlend Temporal Attention (NeurIPS 2024 · similarity 0.85)
- SmartEraser: Remove Anything from Images using Masked-Region Guidance (CVPR 2025 · similarity 0.85)
- Attentive Eraser: Unleashing Diffusion Model’s Object Removal Potential via Self-Attention Redirection Guidance (AAAI 2025 · similarity 0.84)
- ROSE: Remove Objects with Side Effects in Videos (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.