Open-set personalized video generation research record

Freeze date: 2026-08-30 (Asia/Shanghai). This record supports docs/tasks/personalized-video-generation.md. It distinguishes primary-source review, artifact inspection, proposed tests and actual reproduction. No model checkpoint was trained or run in this batch.

1. Review questions

1. When is a reference image a subject definition rather than an I2V time anchor?
2. What does open-set mean when the foundation model's pretraining set is not fully auditable?
3. Which routes adapt one subject at test time, which amortize personalization into an encoder or adapter, and which inject reference tokens without tuning?
4. How do methods bind several references to several entities without identity blending, subject drop or role/action exchange?
5. Which measurements separate identity fidelity, prompt/action following, motion, temporal drift, reference leakage and adaptation cost?
6. What changed from the image-personalization ancestors of 2022–2023 through the formal 2024–2026 video milestones?
7. Which code, weights, data and evaluation artifacts were actually public at the freeze date?

2. Search and source protocol

2.1 Source order

1. CVF, NeurIPS, AAAI, ACM, Springer and OpenReview proceedings;
2. arXiv version pages and paper HTML/PDF for first-public dates and preprints;
3. author project pages, GitHub and Hugging Face for artifact status;
4. no survey, leaderboard screenshot, vendor blog or secondary article was used as evidence for a mechanism or publication status.

2.2 Query families

• subject-to-video open-set personalization multi subject
• video identity customization motion preservation
• tuning-free multi-concept video personalization anchored prompt
• identity preserving video generation DiT frequency decomposition
• video personalization preference optimization reward
• multi-reference video generation temporal interval identity
• subject video dataset benchmark identity leakage binding
• CVPR ICCV NeurIPS 2025 2026 personalized video

2.3 Inclusion rules

A work enters the main lineage when it changes at least one of the following:

• subject acquisition or adaptation surface;
• text–reference or reference–slot binding;
• identity–motion conflict handling;
• multi-subject composition or role/relation control;
• subject-to-video data, evaluation or temporal-presence protocol;
• post-training objective that directly targets personalization failures.

Image-only personalization is retained only as an ancestor. Strict I2V, V2V editing, pose/audio animation, style-only LoRA and motion-only customization are kept as neighboring tasks unless a paper explicitly evaluates subject identity under novel video content.

3. Frozen task boundary

The chapter uses the following operational definition:

Given one or more subject reference sets and a text prompt, optionally with motion, camera, relation or temporal-presence controls, generate a new video in which each intended subject retains its identity and key attributes. The references do not have to occupy the output timeline, and there is no complete source video whose timeline must be preserved.

Consequences:

• a first frame that must be reproduced is strict I2V;
• a full source video that must be edited is V2V;
• a portrait plus pose/audio/driving signal is character or digital-human animation unless the work also studies open-context subject generation;
• per-subject tuning remains personalization, but its data, trainable parameters, steps, time, memory and stored state must be reported;
• open-set means generalization relative to a declared identity- or subject-disjoint evaluation split. It does not establish that an internet-pretrained foundation model has never seen a celebrity, product or object, and it does not mean open weights, open vocabulary or zero-shot.

4. Route taxonomy used in the chapter

Route	Subject-specific work at deployment	Typical mechanism	Principal risk
A. Per-subject tuning	token, LoRA, adapter or full tuning	bind an identifier to a few images	reconstructive overfit, static/copy shortcut, storage
B. Amortized encoder/adapter	none for a new subject	shared vision/face encoder and learned injection	domain bias, insufficient fine detail, hidden training scale
C. Frozen or in-context reference tokens	none or lightweight inference logic	concatenate/reference cross-attend to image tokens	prompt domination, order ambiguity, reference leakage
D. Image personalization plus motion module	separate reusable modules	combine spatial identity and temporal motion modules	module interference and motion forgetting
E. Multi-subject/relation composition	none or several subject modules	masks, anchored prompts, subject slots, relation/motion adapters	blend, drop, attribute/action swap
F. Reward/preference post-training	per-identity or global	identity, semantic, motion or VLM rewards	evaluator gaming, reward overfit, proxy blindness
G. Temporal/multi-shot memory	usually amortized	reference time windows, content anchors, cross-shot context	boundary softness, long-horizon drift, context conflicts

These are alternatives and combinations, not a single upgrade chain.

5. Primary evidence ledger

5.1 Early route-forming work

Work	Safe mechanism claim	Evidence boundary
Textual Inversion / DreamBooth	established learned tokens and per-subject image-model tuning	image generation ancestors, not video evidence
VideoDreamer	early multi-subject language-video customization preprint	no formal proceedings or released method code at freeze date
DreamVideo	3–5 subject images plus motion videos; separate textual identity, identity adapter and motion adapter	authors' 20-subject/30-motion test set; first-attempt language is an author claim
VideoBooth	feed-forward image-prompt video: coarse CLIP embedding plus multi-scale image-latent attention injection	image-level CLIP/DINO; not a strict multi-subject binding protocol
CustomVideo	co-occurrence images, subject masks and attention control for two/three subjects	first public in 2024; formal TMM publication is 2026
DisenStudio	spatial-disentangled cross-attention and masked/motion-preserved tuning	ACM MM 2024; custom training rather than open-set amortized inference
Magic-Me	identity tokens, 3D Gaussian noise prior, face/tiled refinement	ECCV 2024 workshop proceedings, not main ECCV; human identity focus

5.2 2025 formal system milestones

Work	Safe mechanism claim	Critical evidence boundary
ConsisID	CogVideoX DiT with global keypoint/face latent and local recognition+CLIP features; no per-person tuning	single human identity; 30-person benchmark; FID is worse than ID-Animator in its main table
Video Alchemist	entity word + reference image binding, separate personalization cross-attention, no test-time optimization	architecture supports multi-subject, but main quantitative subsets are 1,736 single-subject and 1,285 single-face videos
Movie Weaver	[R1]/[R2] anchored prompts plus concept embeddings encode text–image association and reference order	only five trained configurations; automatic comparison is 97 single-face cases; multi-concept evidence is mainly human ablation and qualitative comparison
VideoMage	spatial subject LoRAs, temporal motion LoRA, auxiliary-video regularization, negative guidance and collaborative sampling	six motions × three subject pairs × four backgrounds; not open-set amortized personalization
PersonalVideo	per-identity video reward tuning, semantic consistency to the base model and late-step identity injection	explicitly single identity and per-identity optimization; no multi-identity support
MagicMirror	face identity/structure dual branch plus cross-modal adapter and two-stage training	human face specialization; formal ICCV 2025
MagicID	identity/dynamics preference pairs and hybrid sampling	identity-focused human customization; reward quality is a dependency
Phantom	joint text-image-video triplets and cross-modal alignment for single/multi-subject S2V	foundation-model scale and data make route-level comparison difficult
DreamRelation	learns a relation from exemplar videos with relational decoupling and dynamics enhancement	relation customization is not identical to open-set subject acquisition
DualReal	alternates identity/motion adaptation and controls both by denoising stage and DiT depth	joint tuning route; its reported relative gains are protocol-specific
OpenS2V-Nexus	OpenS2V-Eval plus OpenS2V-5M subject-text-video triples	automatic metrics remain proxies; data provenance and foundation-model contamination need separate audits

5.3 2026 formal and frontier evidence

Work	Safe mechanism claim	Critical evidence boundary
AlcheMinT	VAE reference tokens, entity text binding and interval-conditioned positional bias for subject appearance/disappearance	tables test one/two references; benchmark/training sample counts are not disclosed; timing and CLIP identity trade off
ID-Crafter	hierarchical intra/inter-subject and cross-modal attention, VLM semantic guidance and online RL	formal CVPR 2026; VLM/reward correctness must be audited separately
Gloria	compact character content anchors, extra/intra-clip cues against copying and weak positional offsets between anchors	authors report videos beyond 10 minutes; this is not an independent reproduction or arbitrary-character guarantee
SMRABooth	aligns self-supervised subject representation and optical-flow motion representation with sparsely injected LoRAs	subject+motion customization; requires per-concept tuning
ConsID-Gen	strict I2V neighbor: auxiliary unposed views and visual-geometric encoder for view-consistent identity	first image is a timeline anchor, so it does not redefine open-set S2V
IPRO	strict human I2V neighbor: differentiates a face reward through late sampling steps with KL regularization	reward improvement does not validate generic objects or non-face identity
ID-Sim	identity-focused feed-forward similarity metric trained for human-selective sensitivity	image/crop metric; no temporal or role-binding measurement by itself
VGBE 2026 challenge	consistent I2V evaluation with identity, geometry and perceptual quality	workshop challenge and I2V contract, not an S2V benchmark
Vera	million-pair cross-clip identity-aligned human data, masked identity supervision and layer-wise reference attention	2026-07 preprint; human-specific and not formally published at freeze date
Keyframe-Anchored	training-free keyframe sequence plus multi-reference interpolation for timestamped actions	challenge solution ranked third; sequential I2V/keyframe route rather than general S2V evidence

6. Red-team checks against overclaiming

6.1 Video Alchemist

• The paper's training reference and target are sampled from the same video. It explicitly documents pose, lighting, occlusion, crop and camera-view leakage, and uses blur, scaling, color, brightness, flip, shear and rotation to reduce, not eliminate, the shortcut.
• MSRVTT-Personalization has 2,130 clips and subject-level metrics, but the main comparisons use only the single-subject and single-face subsets. Multi-subject support is therefore an architectural and qualitative claim, not a complete role-binding benchmark result.
• The paper reports that more references can reduce text alignment. Its own augmentation ablation also shows a Pareto movement rather than improvement on every metric.
• The authors list residual reference-pose/expression copying, segmentation burden, unrealistic multi-subject scale/composition and absence of an automatic visual-quality metric as limitations.

6.2 Movie Weaver

• Training covers face, face-body, face-body-animal, two-face and two-face-body; a three-person prompt can collapse to two because that configuration is absent from training.
• The 228K pretraining videos and 651 curated fine-tuning videos are proprietary Shutterstock-derived data; the final model is 30B and the reported pretraining used 256 H100 GPUs for about five days.
• Its only automatic matched comparison is on 97 single-face cases. The multi-concept claim relies on qualitative comparison to Vidu 1.5 and a 300-pair two-face human ablation.
• Anchored prompts and concept embeddings strongly improve two-face separation, but the protocol does not measure which person performs which action over time. The paper also documents big-face, reduced-motion and unseen-template failures.

6.3 PersonalVideo

• This is per-identity reward tuning: 800 AnimateDiff iterations or 4,000 HunyuanVideo steps in the reported settings, not tuning-free open-set inference.
• The evaluation has 20 identities × 50 prompts = 1,000 videos. The method reports a joint increase in face similarity, dynamic degree and text score, but Dynamic Degree is motion magnitude, not action correctness.
• The paper explicitly states that it cannot generate multiple identities.
• The prose says it minimizes cosine similarity while the displayed identity objective is a positive cosine term in the total loss. Without code, the sign convention is ambiguous; the chapter therefore explains the intended reward without reproducing the questionable equation.

6.4 ConsisID

• The internal pipeline filters to 130K clips; the benchmark uses 30 people, five images each and 90 prompts. “Not in our training set” does not establish absence from CogVideoX or face-recognition pretraining.
• Its main table reports higher face/text scores than ID-Animator, but worse face-region FID (151.82 versus 117.46, where lower is better). The paper's statement that it wins across all metrics is inconsistent with the table.
• The supplement itself notes weak alignment between FID and human perception. The chapter reports metric directions rather than repeating the global claim.
• Loss symbols/names are reversed once in the setup section. The chapter uses full loss names and does not reuse those symbols.

6.5 AlcheMinT

• The main table evaluates one and two references. “Up to 15 tracks per training video” describes the annotation pipeline, not verified 15-reference generation.
• WeRoPE improves the reported temporal interval metrics but lowers both CLIP text and CLIP reference scores in the central ablation; it is a Pareto trade, not a free identity improvement.
• Standard RoPE provides relative phase structure, not a theorem that attention monotonically decays with distance. The paper's rotation-mixture equality and one right-interval timestamp also appear algebraically ambiguous. The chapter therefore describes WeRoPE as an empirical interval-dependent phase bias and does not present the disputed equality as fact.
• S2VTime uses synthetic T2I references and GroundingDINO+SAM2 for interval extraction; it does not test same-class identity swaps or full role/action binding.

7. Evaluation contract derived from the review

The chapter separates these measurements:

1. Subject presence: the intended subject is detected/tracked in expected frames; a missing subject gets zero rather than being omitted.
2. Identity/attribute fidelity: face or generic-subject similarity is computed on tracked subject crops, with per-frame curves and real-video calibration.
3. Temporal drift and re-entry: identity before/after occlusion or leaving the frame is compared directly.
4. Text/action/camera following: global semantics and subject-specific roles are scored separately.
5. Motion: optical-flow magnitude, trajectory diversity and action success are separate; temporal consistency cannot substitute for motion.
6. Binding: the full reference-to-generated-subject similarity matrix is measured; diagonal margin and swap tests reveal blending and role exchange.
7. Leakage: reference background, pose, crop and pixel-copy similarity are explicitly probed under counterfactual references.
8. Base-capability retention: the same prompts without personalization are compared to the frozen base model.
9. Adaptation/system cost: data, steps, trainable/stored parameters, wall time, peak VRAM, generation latency and failure rate are reported.
10. Rights and contamination: consent, license, biometric processing, near-duplicate search, revocation and model-state deletion are audited.

8. Artifact status at freeze date

Work	Code / weights / data status checked
Magic-Me	official training/inference code and some identity embeddings; external base weights required; no training dataset
PersonalVideo	author repository remains a Code Coming Soon placeholder; no weights/data
Video Alchemist	official MSRVTT-Personalization benchmark/evaluation repository; no model code or weights
Movie Weaver	project page; no linked official code, weights or data
MAGREF	inference code and checkpoint; 480P/14B Pro and training code remain TODO; no training data
AlcheMinT	README/project placeholder; no method code, weights or downloadable benchmark
PoCo	project/demo repository; core model code, weights and data not released
OpenS2V-Nexus	evaluation code, generated comparison videos and released dataset subsets under the repository's stated terms
ConsisID	official code/model ecosystem and OpenS2V links; the full internal 130K training corpus is not equivalent to an open dataset

Artifact availability is not model-quality evidence, and a project page is not counted as code release.

9. Teaching visual record

9.1 Final asset

• File: assets/diagrams/personalized-video-binding-contract.png
• Built-in image generation tool; two text/semantics corrections followed by a single-panel failure correction.
• Pixel dimensions: 1672 × 941 (approximately 16:9)
• Color space: RGB
• File size: 1,991,841 bytes
• SHA-256: 856761f1d0fc89c76b90aef4407879d59ab54a28a8d8c1ebcc9161e2d294d5b9

9.2 Final generation/edit intent

The final image depicts three reference sets (human, dog, toy), three independent subject slots, four parallel route choices (TUNE, ADAPTER, IN-CONTEXT, BIND), a generated video in a novel context, five independent gates and four failure panels (BLEND, DROP, COPY, FREEZE). The reference images are visually outside the output timeline.

The base prompt requested an original vector-like scientific infographic with exact uppercase labels, no paper/model names, scores, logos, watermark or copied academic layout. The first correction changed the route column to four true alternatives and changed LEAKAGE to NO LEAKAGE. The second correction changed only the BLEND panel into a visible human–dog–toy hybrid.

9.3 Visual inspection

• [x] All exact labels are present and correctly spelled.
• [x] Route boxes are parallel alternatives, not a sequential upgrade chain.
• [x] Reference sets are not depicted as first frames.
• [x] Human, dog and toy remain separately bound in the successful output.
• [x] Blend, drop, reference-copy and static-freeze failures are distinguishable.
• [x] Color is backed by border pattern, icon and line-style differences.
• [x] Original-size inspection passed.
• [x] Generic Gray profile rendering remains readable; arrows, boxes, subject groups and failure panels do not rely on hue alone.
• [x] No paper figure, model logo, score or watermark appears.

The chapter also includes two editable Mermaid diagrams with accTitle and accDescr: a task-boundary decision tree and a multi-subject binding/falsification graph. Prose alternatives are included for narrow screens and screen readers.

10. Proposed experiment and reproduction boundary

PersonaBind-1 in the chapter is a proposed falsification protocol, not a completed benchmark run. It freezes reference sets, identity-disjoint splits, prompts, seeds, adaptation budgets and evaluators, and changes only the personalization route. No checkpoint was downloaded or executed here, so this batch can claim a primary-source-grounded review, artifact audit and executable evaluation design, but not independent confirmation of any paper's quality, speed, memory, open-set generalization or safety.

11. Validation record

Final integration checks were run on 2026-08-30:

Check	Result
Markdown	markdownlint-cli2 0.23.2 / markdownlint 0.41.1: the 13 non-README changed/new Markdown files passed with 0 issues; README retains exactly two pre-existing MD001/MD028 findings outside the edited lines
Local links and images	307 relative references across 14 Markdown files resolve to 143 unique local targets; 0 missing
Reference anchors	The new chapter has 35 references, 86 citation occurrences and all 35 references cited; no missing, orphan, duplicate, mismatched or skipped anchor
External primary sources	All 35 chapter URLs were checked: 33 returned HTTP 200, the TMM DOI returned 202, and the ACM DOI's automation 403 was separately verified through DOI/Crossref metadata; no deterministic 404, 410, 5xx or timeout
Mermaid	20 blocks across 11 affected documentation pages all contain accTitle / accDescr and rendered with Mermaid CLI 11.16.0 plus system Chrome to non-empty SVG artifacts
Mermaid visual check	Both chapter diagrams and the revised eight-branch reading route were rendered to PNG and inspected in color and grayscale; no clipped node, broken branch or color-only meaning was found
Generated teaching visual	The 1672×941 RGB PNG was re-inspected at original size and in grayscale after integration; labels, route alternatives, three subject slots and blend/drop/copy/freeze failures remain legible; SHA-256 matches the value in section 9
Timeline preservation	The same 75 HTML image src / alt pairs remain in the same order as HEAD; every alt is non-empty and every local image exists
Bibliography	Registry, metadata, BibTeX and generated index contain the same 132 citekeys; section N contains 8 entries, all 66 unique GitHub URLs match the Star snapshot, Python compile and updater --check passed
Patch, credential and size hygiene	git diff --check passed; strong credential-pattern scanning over all 20 changed/untracked files found no candidate; the largest file is the 1,991,841-byte teaching PNG and no changed file exceeds 5 MiB

Temporary Mermaid renderings and grayscale derivatives remain outside the repository. These checks validate the review, source boundaries and teaching artifacts; they do not train, download or independently run a video-generation checkpoint. PersonaBind-1 remains a proposed protocol.