ReMake: Structured Revision Across Modalities

• ReMake is a multidisciplinary paradigm for revising existing artifacts through constrained transformations that preserve essential structures across languages, videos, graphics, and data.
• It employs methods such as retrieval augmentation, localized editing, and targeted re-curation, achieving high accuracy metrics (e.g., over 90% retrieval accuracy in multilingual tasks and improved video consistency scores).
• Practical implementations include multilingual knowledge editing, local video regeneration, and dataset re-curation, demonstrating significant performance improvements over prior methods.

Searching arXiv for the specific ReMaKE paper and closely related "remake" usages to ground the entry in current arXiv records. In current arXiv usage, “ReMake” and cognate forms such as “ReMaKE,” “Remake,” and “re-makeup” do not denote a single method or benchmark. They instead name a family of operations that revise, normalize, reconstruct, or selectively re-curate an existing object rather than generating it ex nihilo. In LLMs, ReMaKE is a retrieval-augmented framework for multilingual knowledge editing; in video generation, ReVideo formulates “remake a video” as local editing with joint content and motion control; in 3D face graphics, BareSkinNet treats de-makeup and de-lighting as the prerequisite for re-makeup; and in dataset construction, SEED produces the compact corpus Honeybee-Remake-SEED-200K through weighted independent set selection (Wang et al., 2023, Mou et al., 2024, Yang et al., 2022, Zhang et al., 15 May 2026). Other uses are broader but structurally related, including slide remaking in digital pathology, screen-remake attacks in face anti-spoofing, the “remake” of a mathematical characterization, and even a modeled electoral rematch (Zhang et al., 2018, Liu et al., 2023, Gounoue, 2020, Crokidakis et al., 2021). This suggests that, across fields, “ReMake” functions as a constrained transformation paradigm: the target is edited, normalized, or re-instantiated while preserving selected structure from an antecedent object.

1. Research scope and recurring meanings

The term appears in several technically distinct senses. In some papers it is an acronym, as in Retrieval-augmented Multilingual Knowledge Editor (ReMaKE); in others it is a task label, as in “Remake a Video”; in others it is an ordinary noun or verb, as in remake the slide, screen remake attacks, break and remake of contacts, or a 2022 remake (Wang et al., 2023, Mou et al., 2024, Zhang et al., 2018, Liu et al., 2023, Lemrich et al., 2017, Crokidakis et al., 2021). The shared research pattern is that an existing artifact is not discarded but reworked under explicit constraints.

Domain	Meaning of “ReMake” or “remake”	Representative source
Multilingual LLM editing	Retrieval-augmented prompt-based knowledge editing across languages	ReMaKE (Wang et al., 2023)
Video generation	Local video editing with joint content and motion control	ReVideo (Mou et al., 2024)
Face graphics	De-makeup and de-lighting as a prerequisite for re-makeup	BareSkinNet (Yang et al., 2022)
Data curation	Task-targeted dataset re-curation	SEED / Honeybee-Remake-SEED-200K (Zhang et al., 15 May 2026)
Digital pathology	Remake or manually review low-quality slides	SlideNet (Zhang et al., 2018)
Face anti-spoofing	Screen replay or screen-remake attack category	CLIPC8 (Liu et al., 2023)

This range matters because the term is not merely rhetorical. In each case it identifies a specific intervention point in a workflow: post hoc factual correction in LLMs, localized regeneration in video, normalization before cosmetic transfer in 3D graphics, compact dataset reconstruction for instruction tuning, or quality-gated re-preparation in microscopy. The term therefore marks a shift from one-shot generation toward iterative, structure-preserving revision.

2. ReMaKE as multilingual knowledge editing

In the most explicit acronymic usage, ReMaKE stands for Retrieval-augmented Multilingual Knowledge Editor, a model-agnostic method for updating LLM behavior when knowledge is edited in one language and queried in another (Wang et al., 2023). The central problem is multilingual knowledge editing: if a counterfactual fact is supplied in language $l_2$, the model should answer correctly when queried in language $l_1$. The paper formalizes retrieval over a multilingual knowledge base $K_{12}=\{k_1^{l_2},\dots,k_n^{l_2}\}$ by

$$k_{i^*}^{l_2}=R(x^{l_1})= \begin{cases} k_{i^*}^{l_2}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=1\ \text{None}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=0 \end{cases}$$

with

$$i^*=\arg\max_i P\big(f_e(x^{l_1},k_i^{l_2})=1\big).$$

The architecture has two stages. First, a multilingual retriever, implemented with a sentence-transformer/XLM-R-based encoder and a binary classification head, maps the query and knowledge items into a shared multilingual embedding space. Second, the retrieved fact is concatenated with the user prompt, optionally together with bilingual few-shot demonstrations, so that the LLM performs in-context editing rather than parameter modification. Zero-shot ReMaKE models

$P(y^{l_1}\mid x^{l_1},k^{l_2}),$

whereas few-shot ReMaKE models

$P(y^{l_1}\mid x^{l_1},k^{l_2},S).$

The bilingual examples are selected by unsupervised cosine-similarity retrieval with all-MiniLM-L6-v2.

The evaluation uses MzsRE, a multilingual version of zsRE covering 12 languages—English, Czech, German, Dutch, Spanish, French, Portuguese, Russian, Thai, Turkish, Vietnamese, and Chinese. After deduplication, the dataset contains 743 unique items and includes paraphrased, unrelated, and portability questions (Wang et al., 2023). The four metrics are Reliability, Generality, Locality, and Portability. The retriever attains over 90% accuracy in all languages and around 97–100% in many language pairs. Empirically, ReMaKE-few-bi is the strongest overall variant, while ReMaKE-zero is usually best for portability because it introduces less prompt interference. On English-edit to other-language testing, the reliability gains over IKE range from about +24.76 for Czech to +58.72 for Russian. At the same time, portability remains low for all methods, and languages such as Chinese, Russian, Thai, and Turkish are harder. A plausible implication is that ReMaKE’s main contribution is not universal reasoning transfer but scalable cross-lingual factual redirection with strong locality.

3. ReMake as local video regeneration with motion and content control

In video generation, the term is operationalized by ReVideo: Remake a Video with Motion and Content Control, which defines local video editing as joint control over what appears and how it moves (Mou et al., 2024). The task takes as input a video $$\mathbf{V}\in\mathbb{R}^{N\times 3\times W\times H}$$ and an editing mask $$\mathbf{M}\in\mathbb{R}^{1\times 1\times W\times H}$$. The unedited content is

$\mathbf{c}_{con}=\mathbf{V}\cdot \mathbf{M},$

content editing is anchored by a modified first frame

$l_1$0

and motion control is given by trajectory maps

$l_1$1

The generator is Stable Video Diffusion, with denoising written as

$l_1$2

under EDM preconditioning.

The core difficulty is a coupling imbalance between dense unedited content and sparse trajectory signals. ReVideo addresses this through a three-stage training strategy. Motion prior training first learns trajectory control in isolation. Decoupling training then constructs masked composites from two different videos,

$l_1$3

so that motion inside the edit region cannot be inferred from the preserved context. Deblocking training finally restores coherence while updating only the key and value embeddings $l_1$4 and $l_1$5 in temporal self-attention layers. The remaining fusion problem is handled by the spatiotemporal adaptive fusion module (SAFM): $l_1$6 This means that content and motion conditions are not added with a fixed weight; the mixture varies across space and denoising time.

On a test set of 16 videos at $l_1$7, ReVideo is compared with InsV2V, AnyV2V, and Pika. The reported human-preference results are 59.1% overall preference and 67.0% editing-target satisfaction for ReVideo, versus 27.9%/23.9% for Pika, 2.8%/4.0% for AnyV2V, and 10.2%/5.1% for InsV2V (Mou et al., 2024). Automatic metrics include CLIP text-alignment 0.2304, temporal consistency 0.9864, and PSNR 32.85. The three canonical applications are local content change with fixed motion, motion change with fixed content, and simultaneous change of both content and motion, and the system also extends to multi-area editing without specific retraining. In this usage, “remake” denotes localized, controllable regeneration rather than whole-video re-synthesis.

4. Visual remaking as normalization and long-domain translation

A second visual sense of ReMake concerns reconstruction across substantial nuisance factors or domain gaps. BareSkinNet is directly relevant to re-makeup because it removes both makeup and illumination while reconstructing a normalized 3D face representation (Yang et al., 2022). The 3D morphable model is

$l_1$8

with full coefficient vector

$l_1$9

and lighting

$K_{12}=\{k_1^{l_2},\dots,k_n^{l_2}\}$0

modeled by spherical harmonics. A pretrained teacher network $K_{12}=\{k_1^{l_2},\dots,k_n^{l_2}\}$1 and trainable student network $K_{12}=\{k_1^{l_2},\dots,k_n^{l_2}\}$2 provide reconstruction, while a U-Net $K_{12}=\{k_1^{l_2},\dots,k_n^{l_2}\}$3 predicts a bare-skin image. The de-makeup loss is

$K_{12}=\{k_1^{l_2},\dots,k_n^{l_2}\}$4

and the full training objective is

$K_{12}=\{k_1^{l_2},\dots,k_n^{l_2}\}$5

with

$K_{12}=\{k_1^{l_2},\dots,k_n^{l_2}\}$6

The key mechanism is that the rendered no-light 3DMM is forced to match the bare-skin prediction, while the rendered lit 3DMM matches the reference image.

Once normalized, the pipeline infers diffuse, normal, roughness, and specular maps by unwarping the bare-skin image, 3DMM diffuse, 3DMM normal, and 3DMM geometry into UV space. The paper states that this is done with a pix2pixHD-style framework, followed by SRGAN upsampling to 4K (Yang et al., 2022). The system therefore supports makeup removal, makeup extraction, makeup transfer or re-makeup, relighting, and 3D makeup avatar creation. It reportedly outperforms BeautyGAN, PSGAN, SCGAN, EleGANt, and LADN, and yields lower RMSE with higher PSNR and SSIM when used as preprocessing for texture reconstruction.

A more radical visual sense appears in “Jurassic World Remake: Bringing Ancient Fossils Back to Life via Zero-Shot Long Image-to-Image Translation”, which defines longI2I as translation across large domain gaps where new visual features and new geometry must be generated (Martin et al., 2023). The benchmark method Revive-2I maps a source image $K_{12}=\{k_1^{l_2},\dots,k_n^{l_2}\}$7 to latent space by

$K_{12}=\{k_1^{l_2},\dots,k_n^{l_2}\}$8

applies partial forward diffusion at levels from 50% to 100%, and decodes under text guidance such as “a photo of the head of a <class name> dog”. On Skull2Dog, the best reported setting is Revive-2I 95%, with FID 143.29, KID 0.0345, All @1 100, and Class @1 92.56, outperforming CycleGAN, ACL-GAN, DDIB, and Null-text inversion (Martin et al., 2023). The paper’s broader claim is that guided diffusion is necessary for large domain gaps because prior knowledge about the target domain is required. Together, BareSkinNet and Revive-2I show two complementary remake regimes: nuisance-disentangling normalization and zero-shot semantic re-instantiation.

5. ReMake as dataset re-curation and targeted selection

In data-centric learning, remake denotes deliberate corpus reconstruction rather than example-level editing. SEED formulates data selection as a Weighted Independent Set (WIS) problem on a similarity graph (Zhang et al., 15 May 2026). With graph $K_{12}=\{k_1^{l_2},\dots,k_n^{l_2}\}$9 and node weights $$k_{i^*}^{l_2}=R(x^{l_1})= \begin{cases} k_{i^*}^{l_2}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=1\ \text{None}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=0 \end{cases}$$0,

$$k_{i^*}^{l_2}=R(x^{l_1})= \begin{cases} k_{i^*}^{l_2}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=1\ \text{None}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=0 \end{cases}$$1

Nodes are weighted by trajectory-based influence and edges encode redundancy. The baseline weight for sample $$k_{i^*}^{l_2}=R(x^{l_1})= \begin{cases} k_{i^*}^{l_2}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=1\ \text{None}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=0 \end{cases}$$2 is

$$k_{i^*}^{l_2}=R(x^{l_1})= \begin{cases} k_{i^*}^{l_2}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=1\ \text{None}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=0 \end{cases}$$3

where the trajectory influence follows TracIn-style gradient alignment.

The two main refinements are node value calibration and local scale normalization. Node calibration restricts influence estimation to a bilateral salient subspace $$k_{i^*}^{l_2}=R(x^{l_1})= \begin{cases} k_{i^*}^{l_2}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=1\ \text{None}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=0 \end{cases}$$4, defined by channels salient on both training and target gradients, so that influence is recomputed as

$$k_{i^*}^{l_2}=R(x^{l_1})= \begin{cases} k_{i^*}^{l_2}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=1\ \text{None}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=0 \end{cases}$$5

Local scale normalization estimates density with the $$k_{i^*}^{l_2}=R(x^{l_1})= \begin{cases} k_{i^*}^{l_2}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=1\ \text{None}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=0 \end{cases}$$6-th nearest neighbor,

$$k_{i^*}^{l_2}=R(x^{l_1})= \begin{cases} k_{i^*}^{l_2}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=1\ \text{None}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=0 \end{cases}$$7

and adds edges only when

$$k_{i^*}^{l_2}=R(x^{l_1})= \begin{cases} k_{i^*}^{l_2}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=1\ \text{None}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=0 \end{cases}$$8

In practice, the conflict graph is built with FAISS, and WIS is solved approximately by a greedy max-heap procedure.

The concrete remake product is Honeybee-Remake-SEED-200K. Starting from Honeybee-1M, annotations are refreshed with Doubao-1.6-VL, SEED is run independently for seven target benchmarks, and examples receive votes

$$k_{i^*}^{l_2}=R(x^{l_1})= \begin{cases} k_{i^*}^{l_2}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=1\ \text{None}, & \text{if } f_e(x^{l_1},k_{i^*}^{l_2})=0 \end{cases}$$9

The top 20% by vote are retained, yielding roughly 200K multimodal instruction pairs (Zhang et al., 15 May 2026). Experimentally, SEED reaches an average score of 57.3 at 5% selection for instruction tuning, matching full-data training and beating TAROT and Nait by about 1.1 points; at 1% it reaches 58.2, exceeding MoNA by 1.0 point. In visual instruction tuning it attains 65.4, above the full-data baseline 65.1 and above DataTailor by 1.2 points. In semantic segmentation it reaches 61.41 mIoU at 20%, which is 3.01 points above full-data training. A highlighted proxy setting uses Qwen3-1.7B to select 2.5% of the data for Phi-4-14B, reducing GPU cost by 2.5× and saving around 138 A100 GPU hours. Here, remake is best understood as principled re-curation: the corpus is rebuilt to optimize quality, diversity, and downstream utility simultaneously.

6. Operational, security, and interactive-media workflows

Outside generative modeling, remake often denotes a decision or threat model inside a larger pipeline. In SlideNet, automatic quality assessment of Gram-stained direct smear slides is explicitly linked to remake decisions (Zhang et al., 2018). If staining quality is low, the system can send a request to the automatic slide preparation system to remake the slide; if damage is severe, it can instead trigger manual microscope reading. The pipeline first performs dense region classification with a MobileNet-based CNN and then derives a region distribution histogram for slide-level judgments. The eight region labels are Crystalized, Damaged, Dense, Dirt, Edge, Empty, EpiOnly, and LeukOnly. A region information-density score is defined by

$$i^*=\arg\max_i P\big(f_e(x^{l_1},k_i^{l_2})=1\big).$$0

with $$i^*=\arg\max_i P\big(f_e(x^{l_1},k_i^{l_2})=1\big).$$1, $$i^*=\arg\max_i P\big(f_e(x^{l_1},k_i^{l_2})=1\big).$$2, and $$i^*=\arg\max_i P\big(f_e(x^{l_1},k_i^{l_2})=1\big).$$3, $$i^*=\arg\max_i P\big(f_e(x^{l_1},k_i^{l_2})=1\big).$$4. On patch classification, SlideNet-224 reaches 86.8% accuracy versus 45.3% for HOG+SVM and 51.2% for colorLBP+SVM, and slide-level agreement with expert opinions is 0.778. In this setting, remake is a quality-gated remediation step.

In face anti-spoofing, CLIPC8 treats screen remake or replay as one element of a richer eight-class taxonomy (Liu et al., 2023). The model uses a pretrained CLIP ViT-L/14 backbone, with image-text contrastive alignment based on cosine similarity

$$i^*=\arg\max_i P\big(f_e(x^{l_1},k_i^{l_2})=1\big).$$5

followed by softmax over eight prompts. Category (f) Recaptured Photo explicitly includes personal photos or ID photos displayed on a mobile phone or computer screen, with cues such as screen borders, moiré patterns, and partial information of the hands. Zero-shot on Replay-Attack, CLIPC8 reports Real 89.03, Fake 85.12, and ACER 0.595%; after broader fine-tuning it reports 99.00 real and 99.81 fake on the same benchmark (Liu et al., 2023). The relevance of remake here is adversarial rather than constructive: the system must detect recaptured or replayed media that visually re-present a prior face sample.

A more narrative use appears in “What if Red Can Talk?”, which studies dynamic filler dialogue generation with knowledge graphs and GPT-4 in Final Fantasy VII Remake and Pokémon (Nananukul et al., 2024). For FFVIIR, the experiments use three bosses, 13 battle situations, and 65 GPT-4 responses. The output is grounded in RDF-style triples for Cloud and boss states, but the paper notes that Cloud often sounds overly positive and that responses are frequently driven more by immediate situation than by explicit knowledge-graph facts. This usage does not define a ReMake framework, but it shows how a title-bearing cultural artifact can become a testbed for constrained regeneration of dialogue.

7. Scientific and mathematical extensions of the remake concept

In several papers, remake is neither a framework name nor an editing task, but a mechanistic or conceptual descriptor. In granular mechanics, dynamic softening is attributed not primarily to microslip but to break and remake of contacts under AC-driven rearrangement (Lemrich et al., 2017). The DEM study defines inter-particle forces with Hertzian normal contact, tangential frictional contact, viscous damping, and rolling resistance, and shows that resonance frequencies shift downward as drive amplitude increases. The Slipping Contact Ratio changes little when friction is varied, whereas the coordination number $$i^*=\arg\max_i P\big(f_e(x^{l_1},k_i^{l_2})=1\big).$$6 decreases at resonance and strongly tracks softening. The paper’s interpretation is that extended structural change—contact breakage and remake—reduces connectivity and hence stiffness.

In political modeling, “After 2018 Bolsonaro victory, is a 2022 remake feasible?” uses remake in the sense of a rematch (Crokidakis et al., 2021). The paper models Brazilian voting trends with left and right groups containing sensitives, inflexibles, and radicals, plus an optional populist group $$i^*=\arg\max_i P\big(f_e(x^{l_1},k_i^{l_2})=1\big).$$7. In the populist scenario,

$$i^*=\arg\max_i P\big(f_e(x^{l_1},k_i^{l_2})=1\big).$$8

for $$i^*=\arg\max_i P\big(f_e(x^{l_1},k_i^{l_2})=1\big).$$9, and when $P(y^{l_1}\mid x^{l_1},k^{l_2}),$0 the model yields $P(y^{l_1}\mid x^{l_1},k^{l_2}),$1 and

$P(y^{l_1}\mid x^{l_1},k^{l_2}),$2

With $P(y^{l_1}\mid x^{l_1},k^{l_2}),$3, this gives $P(y^{l_1}\mid x^{l_1},k^{l_2}),$4. The paper concludes that a 2022 Lula/Bolsonaro remake is feasible in either scenario: Lula can win in the low-populism regime if $P(y^{l_1}\mid x^{l_1},k^{l_2}),$5 shifts sufficiently, whereas Bolsonaro becomes favored if the populist coupling crosses the threshold near $P(y^{l_1}\mid x^{l_1},k^{l_2}),$6.

In analysis, “A Remake of Bourgain-Brezis-Mironescu Characterization of Sobolev Spaces” generalizes the BBM limit formula from classical fractional kernels to concentrated $P(y^{l_1}\mid x^{l_1},k^{l_2}),$7-Lévy kernels (Gounoue, 2020). The paper studies nonlocal energies

$P(y^{l_1}\mid x^{l_1},k^{l_2}),$8

proves that finiteness of the liminf characterizes $P(y^{l_1}\mid x^{l_1},k^{l_2}),$9 for $P(y^{l_1}\mid x^{l_1},k^{l_2},S).$0 and $P(y^{l_1}\mid x^{l_1},k^{l_2},S).$1 for $P(y^{l_1}\mid x^{l_1},k^{l_2},S).$2, and identifies the universal constant

$P(y^{l_1}\mid x^{l_1},k^{l_2},S).$3

On extension domains, the nonlocal characterization becomes necessary and sufficient; without the extension property, the paper establishes weak convergence of the associated Radon measures to the local gradient measure. In this context, remake signifies a reconstruction of a classical theorem in a more general kernel and domain setting.

Taken together, these uses show that “ReMake” is not a singular object in arXiv research culture. It is a recurrent label for technically constrained revision: edited facts must remain local, regenerated video must preserve unedited regions, normalized faces must separate makeup from lighting, re-curated datasets must remain informative and diverse, and even physical, mathematical, or political systems are described through the logic of remaking when structure is reorganized without complete replacement.