VideoEraser Techniques

• VideoEraser is a family of video editing methods designed to remove targeted visual content while maintaining spatio-temporal coherence in dynamic scenes.
• It covers observed-video editing, physics-aware removal, and text-to-video concept erasure using diffusion transformers and language models.
• Advanced systems like EraserDiT and SEDiT leverage full 3D attention and prompt-guided mechanisms to address challenges such as shadows, reflections, and motion.

VideoEraser denotes a family of erasure-oriented video technologies whose common objective is to remove targeted visual content while preserving the remaining video as much as possible. In recent arXiv usage, the name spans two technically distinct settings: observed-video editing, where a system removes an object, subtitle, or text region from an existing video and reconstructs the missing content with spatio-temporal consistency, and text-to-video concept erasure, where a plug-and-play mechanism prevents a generative model from producing videos containing a specified concept even when that concept is explicitly or adversarially prompted (Liu et al., 15 Jun 2025, Xu et al., 21 Aug 2025).

1. Task definitions and scope

In observed-video editing, the canonical formulation uses a video $V \in \mathbb{R}^{T \times H \times W \times 3}$ and a binary mask sequence $M \in \{0,1\}^{T \times H \times W}$, and seeks an inpainted video $\hat V$ such that known pixels are preserved wherever $M_t(i,j)=0$ and masked pixels are replaced by plausible, temporally consistent content wherever $M_t(i,j)=1$. The central difficulties are large masks, long-term temporal consistency, and the need to leverage long-range temporal features under large motion, camera pans, and dynamic scenes. Earlier flow-based propagation and spatio-temporal Transformers are described as struggling with large masked areas, heavy motion, and high resolution (Liu et al., 15 Jun 2025).

A stricter formulation appears in physics-aware video instance removal, where the task is a mapping

$(V, M, P) \mapsto \hat V$

with a natural-language removal prompt $P$, under the additional requirement that object-induced physical derivatives such as shadows, reflections, illumination changes, and occlusion patterns are correctly updated or removed. In this setting, high-quality output requires not only visual plausibility but also physical consistency over 81-frame clips (Li et al., 7 Apr 2026).

A related supervised formulation defines video object removal as a mapping

$$\Phi:(\mathbf{V}_{\mathrm{ori}},\mathbf{M}_{\mathrm{obj}})\longrightarrow \tilde{\mathbf{V}},$$

where “complete and stable” removal means that the object and all object-induced side effects—shadows, reflections, highlight changes, and indirect illumination changes—are corrected while maintaining temporal coherence. This distinguishes semantics-driven object removal from generic video inpainting, since side-effect regions may lie outside the explicit object mask (Liu et al., 2 Apr 2026).

2. Diffusion-transformer VideoEraser systems for object removal

One influential observed-video realization is the system built around EraserDiT. It combines automatic object detection, interactive object selection, automatic prompt generation, mask construction, latent diffusion-based completion, and long-sequence inference. A Video LLM, specifically MiniCPM-o, analyzes the video, detects objects, generates a scene description, and supports a short dialogue in which a prompt describing the scene without the removed object is produced. Masks in the demo system can be created by interactive tools such as SAM2 or Grounded SAM2. The masked video is formed as

$$V_{\text{test}}^{\text{masked}}=V_{\text{test}} \ast M_{\text{test}},$$

and the prompt is encoded with a T5-like text encoder for conditioning of the diffusion transformer (Liu et al., 15 Jun 2025).

Architecturally, EraserDiT is built on LTX-Video and uses a 3D causal VAE with spatio-temporal downsampling $32 \times 32 \times 8$, so that the latent temporal length satisfies

$M \in \{0,1\}^{T \times H \times W}$0

During training, latent features and downsampled masks are concatenated along the channel dimension,

$M \in \{0,1\}^{T \times H \times W}$1

and prompt tokens are fused with video tokens by cross-attention. The method adopts Rectified Flow rather than classical DDPM, while retaining the same high-level diffusion interpretation. Unlike earlier video transformers that separate spatial and temporal attention, it uses full 3D attention so that all tokens across time and space interact in a single self-attention mechanism. Training is reported at $M \in \{0,1\}^{T \times H \times W}$2 resolution and 81 frames, for 280k iterations on 24×A100 GPUs, with Adam-W and initial learning rate $M \in \{0,1\}^{T \times H \times W}$3. At inference, the system uses Circular Position-Shift for longer sequences than the trained sequence length. Without any acceleration method, it takes 180 seconds on one NVIDIA A100 GPU to complete a $M \in \{0,1\}^{T \times H \times W}$4 video with 121 frames (Liu et al., 15 Jun 2025).

Circular Position-Shift is introduced to reduce boundary artifacts and inconsistent global motion under long-sequence sliding-window inference. The method constructs a reflected sequence of length $M \in \{0,1\}^{T \times H \times W}$5, shifts window alignments across denoising steps, and extracts the original latent subsequence at the end. The stated motivation is that each denoising step sees a different window alignment, while reflect padding preserves physically coherent temporal neighborhoods (Liu et al., 15 Jun 2025).

3. Understanding-centric and effect-aware removal

A later line of work treats erasure as an understanding problem rather than only a masked inpainting problem. In “From Understanding to Erasing,” the backbone is Wan2.1, a latent video DiT trained with flow matching. Two guidance mechanisms are added. Externally, Object-Induced Relation Distillation transfers relationships between objects and their induced effects from DINOv2 to the video diffusion model. The side-effect region is defined from paired object-present and object-free videos using

$M \in \{0,1\}^{T \times H \times W}$6

with $M \in \{0,1\}^{T \times H \times W}$7, and

$M \in \{0,1\}^{T \times H \times W}$8

Internally, Framewise Context Cross-Attention conditions each denoising block on CLIP visual features extracted from the unmasked background,

$M \in \{0,1\}^{T \times H \times W}$9

together with an empty text prompt. The total loss adds the OIRD term to the flow-matching objective with $\hat V$0, and the reported experiments apply OIRD at the 15th block of the denoiser (Liu et al., 2 Apr 2026).

The OIRD loss aligns cosine-similarity relations between object tokens and side-effect tokens in the denoiser with those in the DINOv2 feature space: $\hat V$1 This formulation explicitly targets non-local object–effect coupling rather than only reconstruction inside the mask (Liu et al., 2 Apr 2026).

Effect-aware removal is developed further in EffectErase. Its training resource, VOR, contains 60,000 video pairs, 145+ hours of content, 12.5M image frames, 443 different scenes, and 366 distinct object categories, with five effect types: occlusion, shadow, lighting, reflection, and deformation. Each triplet provides a video with object and effects $\hat V$2, a matched clean-background video $\hat V$3, and an object mask sequence $\hat V$4. EffectErase is built on Wan 2.1 with 1.3B parameters, uses latent diffusion with a DiT backbone, and frames removal and insertion as inverse tasks. The conditioning for removal is $\hat V$5, whereas insertion uses $\hat V$6. Task-Aware Region Guidance combines task tokens with CLIP foreground tokens, and Effect Consistency uses KL divergence between attention-derived effect maps and a difference-map prior: $\hat V$7 Training uses 81 consecutive frames and $\hat V$8 input resolution (Fu et al., 19 Mar 2026).

This trajectory suggests a shift from mask-bound completion toward effect-aware recovery of a “world without the object,” including corrections to reflections, shadows, lighting, deformation, and occlusion patterns (Liu et al., 2 Apr 2026, Fu et al., 19 Mar 2026).

4. Subtitle and text erasure

An earlier precursor is Scene Text Eraser, which addresses privacy-sensitive text removal in natural scene images. It uses a feedforward CNN-based encoder–decoder with symmetric skip connections, processes $\hat V$9 patches with stride $M_t(i,j)=0$0, and reconstructs the full image from the central $M_t(i,j)=0$1 region of each output patch. The encoder has 4 convolutional layers and the decoder has 4 deconvolution layers, all with filter size $M_t(i,j)=0$2, stride 2, and padding 1. Skip fusion is defined as

$M_t(i,j)=0$3

and training minimizes the L2 reconstruction loss

$M_t(i,j)=0$4

The implementation uses Caffe, SGD, learning rate $M_t(i,j)=0$5, and ReLU activations. On ICDAR2013, the Erased3 variant reduced ICDAR-evaluation recall, precision, and F-score from 82.56%, 83.70%, and 83.13% on original images to 8.35%, 54.07%, and 14.46%, respectively, and reduced DetEval recall, precision, and F-score from 81.90%, 87.15%, and 84.45% to 8.89%, 54.53%, and 15.30% (Nakamura et al., 2017).

The modern video-specialized counterpart is SEDiT, a mask-free video subtitle erasure method via a one-step diffusion transformer. It is described as mask-free, one-stage, one-step, and streaming. Using LTX-Video as the backbone, it encodes the original subtitle-bearing video to a latent $M_t(i,j)=0$6, forms a noisy latent

$M_t(i,j)=0$7

patchifies both the clean reference latent and the noisy latent, and concatenates them along the sequence dimension with identical 3D RoPE coordinates. The model is trained under rectified flow with

$M_t(i,j)=0$8

and a focal weighting over subtitle regions

$M_t(i,j)=0$9

with $M_t(i,j)=1$0. The paper further states that, under localized editing, the conditional optimal transport map and its induced rectified flow velocity field are Lipschitz continuous with respect to the latent variable, which is used to justify one-step sampling. For temporal consistency, the first frame is replaced by its clean latent with probability $M_t(i,j)=1$1 during training. In inference, shot-level segmentation uses TransNetV2 and chunk sizes are 121 frames for 720p, 65 frames for 1080p, and 41 frames for 1440p, enabling native 1440p processing for videos of infinite length (Hui et al., 14 May 2026).

5. Concept erasure in text-to-video diffusion models

In generative safety literature, VideoEraser refers specifically to a training-free framework for concept erasure in text-to-video diffusion models. Its goal is not to edit an existing video but to prevent a T2V model from generating videos containing an undesirable concept—object, artistic style, celebrity, or explicit content—even when that concept is explicitly or adversarially prompted. The framework is described as plug-and-play and model-agnostic across representative T2V diffusion models. It operates through two stages: Selective Prompt Embedding Adjustment and Adversarial-Resilient Noise Guidance. Reported evaluation covers four tasks—object erasure, artistic style erasure, celebrity erasure, and explicit content erasure—and states that undesirable content is reduced by 46% on average across the four tasks compared to baselines (Xu et al., 21 Aug 2025).

SPEA begins with prompt embeddings $M_t(i,j)=1$2 and erase-concept embeddings $M_t(i,j)=1$3, forms pooled vectors $M_t(i,j)=1$4 and $M_t(i,j)=1$5, and defines projection matrices

$M_t(i,j)=1$6

With $M_t(i,j)=1$7, token-level sensitivity is estimated by masking token $M_t(i,j)=1$8, recomputing the pooled embedding, projecting onto $M_t(i,j)=1$9, and evaluating

$(V, M, P) \mapsto \hat V$0

A token is marked as a trigger when $(V, M, P) \mapsto \hat V$1, with $(V, M, P) \mapsto \hat V$2. Trigger-token embeddings are then selectively replaced using

$(V, M, P) \mapsto \hat V$3

ARNG is described as adding a subtract-ive semantic guidance term that pushes predicted noise away from the target concept’s noise direction, while adapting erasure strength across timesteps and frames to preserve step-to-step and frame-to-frame consistency (Xu et al., 21 Aug 2025).

Related work broadens the notion of generative VideoEraser. ICE is a training-free, modality-agnostic, one-shot weight modification method that constructs erase and preserve subspaces with anisotropic energy-weighted scaling, defines an overlap projector

$(V, M, P) \mapsto \hat V$4

solves a convex Spectral Unlearning Objective in closed form, and applies the resulting dissociation operator as a permanent weight edit with zero inference overhead (Biswas et al., 24 Nov 2025). EraseAnything++ targets rectified-flow and flow-matching transformers, formulates concept erasure as constrained multi-objective optimization, introduces implicit gradient surgery for utility-preserving unlearning, uses LoRA-based parameter tuning with attention-level regularization, and adds an anchor-and-propagate mechanism in the video setting to mitigate temporal drift (Fan et al., 1 Mar 2026).

6. Evaluation, failure modes, and relation to general video editing

The strongest evaluation framework for object-removal VideoEraser systems is the Physics-Aware Video Instance Removal benchmark. PVIR contains 95 high-quality videos at 720p and 81 frames, with instance-accurate masks and removal prompts, and is split into a Simple subset of 57 videos and a Hard subset of 38 videos. Evaluation is decoupled into three human-rated dimensions on a 1–4 scale: Instruction Following, Rendering Quality, and Edit Exclusivity. On the full set, PISCO-Removal achieves IF 3.62, RQ 3.28, EE 3.58, Overall 3.49; UniVideo achieves IF 3.06, RQ 3.45, EE 3.53, Overall 3.35; DiffuEraser achieves IF 2.89, RQ 2.63, EE 3.52, Overall 3.01; and CoCoCo achieves IF 1.60, RQ 1.84, EE 3.07, Overall 2.17. The Hard subset is explicitly designed around reflections, shadows, specular interactions, and stronger physical coupling, and all methods exhibit a persistent performance drop there. Reported failure modes include lingering shadows, broken reflections, incorrect lighting, spatial misalignment, flicker, ghosting, and, for DiffuEraser specifically, spatial blurring and halo artifacts (Li et al., 7 Apr 2026).

A broader editing context is provided by EasyV2V, an instruction-based framework that supports flexible inputs such as video+text, video+mask+text, and video+mask+reference+text. Its model design relies on simple sequence concatenation for conditioning with light LoRA fine-tuning, and its control mechanism unifies spatiotemporal control through a single mask mechanism. In an erasure setting, this makes EasyV2V a general-purpose mask-and-instruction editor rather than a method specialized for physical side-effect removal or safety-oriented concept unlearning (Mai et al., 18 Dec 2025).

Across these lines of work, a consistent theme is that erasure quality is limited not only by fill-in realism but also by how well a system models non-local dependencies: long-range temporal structure in diffusion transformers, object–effect relations in understanding-centric removal, and semantic overlap in concept-erasure methods. The persistent performance drop on physics-aware hard cases suggests that high-quality VideoEraser systems remain constrained by incomplete modeling of reflections, shadows, illumination interactions, and other causal consequences of removing a target from a dynamic scene (Li et al., 7 Apr 2026).