Re:Cognize -- Open-Set Comic Character Re-Identification

Aaditya Baranwal; Madhav Kataria; Yogesh S Rawat; Shruti Vyas

cs.CVSep 27, 2026

Re:Cognize -- Open-Set Comic Character Re-Identification

Authors: Aaditya Baranwal, Madhav Kataria, Yogesh S Rawat, Shruti Vyas

Organizations: Institute of Artificial Intelligence, University of Central Florida, Orlando, FL, USA · University of Central Florida

Abstract

A manga reader meets a character on one page and knows them on sight a hundred pages later, without ever being handed a cast list. Re-identifying comic characters demands the same, open-set and sequential: pages arrive as a stream in reading order, new faces appear before anyone names them, and the cast is assembled as the story is read. $\textbf{Re:Cognize}$ evaluates recognition as the story is read, not against a cast handed over in advance: four protocols on one query stream, from closed-set retrieval to a cast the model must build and grow itself. The surprise is where models fail. Recognising is close to solved: one reference image per character already ranks as well as a gallery built in advance. Knowing what to believe is not: a model that adds its own matches makes its cast worse, while the same growth with correct labels would gain over twenty points of top-1 accuracy. The bottleneck is acceptance, not vision, and one comparison decides it: an addition pays exactly when it is right more often than the cast already was on the queries it takes over. The comparison has nothing to fit, and measured on half of a new corpus it calls the other half correctly. $\textbf{ReCast}$ puts it to work with nothing fitted on data: a cast sheet of one running average per character, grown only where the page itself vouches for a crop. It recovers a third to two thirds of what perfect labels would, depending on whether the cast starts from random examples or from first appearances. Re:Cognize measures whether a model can read along; ReCast is a cast that does. Our claims are on identity maintenance, recognising characters already met; the emergence of new ones is measured as a diagnostic under a fixed reference rule, and we propose no method for it.

Figures & tables

Figure 2 : The four Re:Cognize protocols. All four answer one stream of query crops in reading order and differ in the gallery and whether it may change. $k$ is the number of seed crops per character, $\tau_{\mathrm{nov}}$ the novelty threshold of P3 and $B_{\max}$ the cap on crops P4 adds per character; frames are coloured by character and dashed where the model added the crop. P3 is a diagnostic of emergence, and its fixed threshold is a reference rule rather than a clustering algorithm we propose.

Method	Closed Set	Open Set	Sequential	Online Gallery	Cross Corpus	Memory
TransReID [ 13 ]	✓	–	–	–	✓	–
OSNet [ 55 ]	✓	–	–	–	✓	–
Instruct-ReID [ 14 ]	✓	–	–	–	✓	–
Zhang et al. [ 52 ]	✓	–	✓	–	–	–
Zhang & Chu [ 51 ]	✓	–	–	–	–	–
Soykan et al. [ 37 ]	✓	–	–	–	–	–

Table 1 : Method positioning. Prior comic and person Re-ID against Re:Cognize across the six capabilities its protocols exercise; Memory is the memory block baseline Re:Cognize evaluates. Cross corpus means evaluation on a second annotated corpus of the same medium.

		P1: Closed-Set		P2: Seeded k=1 , mAP		P4: Seq-R k=1 , id. R-1			P4: Seq-T k=1 , id. R-1
Backbone	Config	mAP	R-1	Seq-R	Seq-T	Static	Pred.	Oracle	Static	Pred.	Oracle
Chance	Random ranking	33.1	30.1	33.9	33.9	12.9	–	–	12.9	–	–
TransReID	Finetuned	37.4	40.6	38.6	33.2	17.4	15.0	41.7	12.8	12.6	42.2
	FT + Mem + LoRA †	38.1	39.9	39.0	34.2	18.1	16.1	42.3	13.9	13.3	42.4
MagiV2	Finetuned	51.2	57.2	54.4	52.9	36.8	36.3	59.1	36.6	37.9	59.0
	FT + Mem †	51.7	56.3	54.6	51.3	37.2	35.9	59.8	35.3	36.9	60.4

Table 2 : Cross-protocol summary on POPCharacters , 8 held-out series,

k=1

seed per character, mean over three training runs. Chance is a random ranking of the same galleries. Per backbone the top row, Finetuned , is the frozen backbone with a trained BNNeck and no LoRA, and the bottom row (

\dagger

) the best memory-block configuration by P1 mAP. Seeds are random (Seq-R) or first appearances (Seq-T). P4 is identity Rank-1 at

B_{\max}=50

under three policies: static , the P2 gallery; predicted , adding each query under its top-1 match; oracle , adding it under its true character.

Panel of Figure 3 — Figure 3 : What the protocols measure, and where the headroom is. (a) One seed per character recovers the closed-set mAP but only part of its Rank-1. (b) Growth by the model’s own top-1 matches stays below a static gallery, while correct labels would add over twenty points (labels: oracle minus static). (a, b): memory-block configuration, one random seed per character, three training runs. (c) Encoder-side adaptation helps most on the backbones weakest in this domain. (d) Each point is one change to the gallery on one backbone and corpus: adding under the page constraint (Section 6 ), adding by top-1 match, or restricting the candidate characters. The stronger the gallery already is, the less any change adds, and Section 5 predicts which points fall below zero.

Backbone	Rule	#clusters	Purity	Hung. Acc	ARI	NMI
TransReID	fixed $\tau_{\mathrm{nov}}=0.55$	95.8	60.2	15.1	2.0	21.8
MagiV2	fixed $\tau_{\mathrm{nov}}=0.55$	38.9	69.6	46.6	24.0	34.1
TransReID	fixed $\tau_{\mathrm{nov}}=0.80$ (loose)	431.1	93.0	4.8	0.2	37.0
MagiV2	fixed $\tau_{\mathrm{nov}}=0.80$ (loose)	190.9	84.3	24.6	9.8	38.5

Table 3 : P3 identity emergence on full per-series streams ( POPCharacters test series, finetuned backbones from the first training run, macro over 8 series). The fixed rule at

\tau_{\mathrm{nov}}=0.55

is the reference instantiation. The loose rows show why Purity cannot be read alone: raising the threshold multiplies clusters four- to fivefold and Purity rises with them, while Hungarian accuracy and ARI collapse. Alternative decision rules are in Appendix Table 15 and the comparison with the memory block in Appendix Table 16 ; neither is a baseline of the framework.

Corpus	Backbone	$\boldsymbol{a}$	$\boldsymbol{c}$	$\boldsymbol{p_{\mathrm{eff}}}$	$\boldsymbol{a^{+}}$	$\boldsymbol{c(p_{\mathrm{eff}}\!-\!a^{+})}$	measured
POPCharacters	TransReID	23.82	34.96%	31.532	23.791	+2.71	+2.71
POPCharacters	InstructReID	25.38	33.15%	30.851	23.437	+2.46	+2.46
POPCharacters	ReID5o	28.68	35.14%	36.682	28.106	+3.01	+3.01
POPCharacters	MagiV3	32.88	35.67%	39.119	31.918	+2.57	+2.57
Manga109	MagiV3	34.03	43.35%	39.513	33.519	+2.60	+2.60
POPCharacters	MagiV2	44.35	34.67%	47.832	47.960	-0.04	-0.04

Table 4 : The commit condition reproduces every measured change. Commitment under the page constraint (Section 6 ) on the bag of

k=5

random seeds per character, without the cast sheet, over three seed draws.

a

is the static gallery’s identity Rank-1 and

c

p_{\mathrm{eff}}

and

a^{+}

are the terms of Equation 1 , in percent. Weighting

p_{\mathrm{eff}}

and

a^{+}

by each record’s capture rate,

\sum_{i}c_{i}x_{i}/\sum_{i}c_{i}

, makes

c\,(p_{\mathrm{eff}}-a^{+})

exactly the mean measured change, so the last two columns agree.

Panel of Figure 4 — Figure 4 : Re:Cast on Re:Zero crops from Re:Verse [ 5 ] , with Rom as identity A. Top: the three changes of Section 6 . A page names a character when one of its crops is already committed to it, and commitment then adds that crop’s page-group sibling. Bottom: over six pages the cast sheet is updated only on the four that name Rom; elsewhere the rule abstains.

Backbone	Static	Cast sheet	$+$ commitment	$+$ expansion	Oracle
POPCharacters , 8 test series, Seq-R, $k=1$
TransReID	17.17	+0.00		+4.71	—
MagiV2	35.95	+0.00		+0.62	—
MagiV3	21.79	+0.00		+4.84	—
InstructReID	16.83	+0.00		+6.84	—
ReID5o	21.03	+0.00		+5.66	—

Table 5 : Re:Cast , one change at a time : P4 identity Rank-1 with random seeds at

B_{\max}=50

, on the 8 held-out POPCharacters series and, zero-shot, on 27 held-out Manga109 volumes. Each column after Static adds one change and gives the cumulative gain over Static , except

+

expansion , which is scored against its own static reference with the expanded crops removed from the queries; Oracle adds every query under its true label. Cells are one training run and three seed draws, and bold marks

p<0.05

on a paired

t

-test over series. At

k=1

the first two changes coincide, the oracle is in Table 2 , and seed expansion was not run on Manga109.

Panel of Figure 5 — Figure 5 : Binding, and pricing the binder. Top, on Re:Verse’s Re:Zero annotations [ 5 ] : (a) a page’s crops are grouped above one similarity threshold, (b) page groups merge in reading order into the best earlier match above a second, or start new ones, and (c) a group’s first seed names it. Bottom: any frozen encoder can bind, with threshold $\tau$ ; its crops are committed when Equation 1 predicts a positive $\Delta$ , with $p_{\mathrm{eff}}$ the additions’ precision on the queries they capture and $a^{+}$ the static gallery’s accuracy there, and $\tau$ is chosen by that prediction, never the measured gain.

Figure 6 : Where the supervision sits. Labelled share of the last twenty crops read, by quarter of the stream, over the 8 held-out POPCharacters series at

k=5

. Both seedings place the same number of labels in the volume and distribute them differently.

		Shared binder		Self		Controls
Backbone	$a$	POPCharacters	Manga109	POPCharacters	Oracle	Shuffled	Seq-R
TransReID	20.7	+14.32	+10.32	+4.96	+20.94	-5.56	+3.93
InstructReID	20.4	+15.45	+10.34	+4.60	+22.46	-5.93	+3.86
ReID5o	20.9	+16.94	+10.35	+2.52	+24.63	-6.07	+2.70
MagiV3	26.5	+16.45	+8.00	+1.73	+23.56	-10.71	+1.89
MagiV2	37.4	+12.51	-7.44	+8.46	+21.25	-20.44	-4.92

Table 6 : Two-stage binding under chronological seeding : change in P4 identity Rank-1 over the static gallery (

a

on POPCharacters ) at

k=5

, averaged over series and seed draws; bold is a gain at

p<0.05

paired over series. Columns are POPCharacters unless marked Manga109 (27 volumes, zero-shot). Shared is the binder of Section 7 for every backbone and Self each backbone’s own. Oracle is the same representation’s ceiling (

+20.2

+26.4

on Manga109), Shuffled permutes the added identities (

-8.1

-43.5

on Manga109), and Seq-R the same binding with random seeds.

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7 : The memory block. A frozen backbone produces L2-normalised features via BNNeck. Working Memory (cross-attention over recent observations) and Episodic Memory (cross-attention over identity prototypes) produce residuals

\delta^{wm}_{t},\delta^{em}_{t}

, combined by Gated Fusion through a small-init projection. Optional LoRA adapts the last

L

backbone attention layers. Trainable: MLPs and attention projections; frozen: backbone

\mathcal{B}

; runtime state: prototypes

\mathbf{P}_{c}

and FIFO buffers

\mathbf{B}_{c}

Configuration	Backbone	BNNeck	Memory	LoRA
Pretrained	frozen	–	–	–
Finetuned	frozen	trained	–	–
Finetuned + Memory	frozen	trained	trained	–
Finetuned + LoRA	adapted	trained	–	trained
Finetuned + Memory + LoRA	adapted	trained	trained	trained

Appendix

Table 7 : Five configurations analysed per backbone: BNNeck fine-tuning, the memory block, and LoRA toggled along three axes.

Backbone	Architecture	Pre-training	Input	$d$
TransReID [ 13 ]	ViT-B/16	ImageNet + Re-ID	$256\!\times\!128$	768
MagiV2	ViT-B	Manga embeddings	$224\!\times\!224$	768
MagiV3	Florence-2 [ 50 ]	Manga comprehension	$384\!\times\!384$	1024
InstructReID [ 14 ]	ViT-B/16	Multi-modal Re-ID	$256\!\times\!128$	768
ReID5o	CLIP ViT-B/16 [ 29 ]	CLIP + Re-ID	$384\!\times\!128$	512

Appendix

Table 8 : Backbone architectures evaluated by Re:Cognize .

d

is the native feature width, at which each backbone’s BNNeck and memory block operate.

Figure 8 : Two-pass open-set inference schematic. Pass 1 produces a coarse identity hypothesis. Pass 2 uses the hypothesis to route Working Memory and emits the final descriptor. The FIFO buffer and prototype bank update only after Pass 2.

	Latency (ms/crop)			Parameters
Backbone	Backbone	$+$ Memory	Block alone	Backbone	Memory block
TransReID	5.8	17.8	4.9	85.6M	10.1M
MagiV2	6.8	20.2	5.0	85.8M	10.1M
MagiV3	24.2	59.6	4.8	360.7M	17.9M
InstructReID	7.2	20.3	4.9	85.8M	10.1M
ReID5o	13.4	34.3	4.8	79.1M	4.5M

Appendix

Table 9 : Computational overhead on one NVIDIA RTX A6000 at batch size one in FP32, median over 300 crops of one test series.

+

Memory is the two-pass inference as implemented, which runs the backbone once per pass, and block alone is the same two passes over cached backbone features. Backbone parameters count what the crop forward touches, since MagiV3 and ReID5o ship modules it never calls. The block’s buffers and prototypes are filled at test time and are not parameters.

Backbone	Configuration	mAP	R-1	R-5	R-10
TransReID	Pretrained	37.1	39.7	79.0	89.2
	Finetuned	37.4	40.6	78.9	89.0
	Finetuned + Memory	37.5	39.5	77.9	88.5
	Finetuned + LoRA	37.7	41.1	78.9	89.0
	Finetuned + Memory + LoRA	38.1	39.9	78.1	88.4
MagiV2	Pretrained	50.8	57.1	85.1	91.0

Appendix

Table 10 : Full P1 closed-set retrieval on the 8 held-out POPCharacters series (Appendix A.16 ). Best per backbone in bold , second-best underlined . Finetuned trains a BNNeck on a frozen backbone; only LoRA rows adapt backbone weights. Every row is the mean over three training runs.

		Finetuned		FT + Memory		FT + Mem + LoRA
Manga	#C	mAP	R-1	mAP	R-1	mAP	R-1
Bakuman	7	62.9	70.2	63.5	68.7	63.5	70.2
Demon Slayer Kimetsu No Yaiba	11	47.2	52.7	48.5	52.5	48.0	51.7
Dr Stone	4	63.8	65.6	63.8	65.0	63.8	66.0
Hunter X Hunter	12	40.7	48.8	41.2	46.5	41.6	45.4
Kagurabachi	6	41.2	51.0	42.0	52.4	41.7	52.3

Appendix

Table 11 : Per-manga P1 results for MagiV2 across three configurations, mean over three training runs and five seed draws. #C is the character count. Best per series in bold .

	Manga109 , 27 volumes				Re:Verse , 1 series, pretrained
	P1 mAP			P4 id. R-1	P1
Backbone	Pretrained	Finetuned	$+$ Mem.	$+$ Mem.	mAP	R-1
TransReID	28.4	29.1	30.3	10.7	34.5	46.0
MagiV2	65.3	65.9	67.4	47.1	84.2	91.0
MagiV3	37.4	39.4	41.5	18.4	48.2	73.0
InstructReID	28.6	30.4	35.5	14.3	30.7	50.9

Appendix

Table 12: Transfer beyond POPCharacters , headline rows. Manga109 is 27 held-out volumes. Pretrained is the released backbone with no POPCharacters training at all; Finetuned and

+

Mem. are POPCharacters -trained checkpoints scored zero-shot here from one training run, the latter the memory-block configuration with the highest P1 mAP on this corpus, so the first gap is what BNNeck fine-tuning on another corpus is worth and the second is the memory block. P4 is identity Rank-1 at

k=1

under random seeding at

B_{\max}=50

, for that same configuration. Re:Verse is one series (Re:Zero) and its two columns are the released backbone with no POPCharacters training. Manga109 cells are the mean over three seed draws, Re:Verse cells over five. The full tables are Table 13 and Table 14 in the appendix. Three vision-language models on the same Re:Verse benchmark reach 1.11, 0.00 and 0.85 percent character-identification accuracy [ 5 ] .

		P1: Closed-Set		P2-R k=1	Seq-R k=1 , id. R-1
Backbone	Config	mAP	R-1	mAP	P2	P4
TransReID	Pretrained	28.4	35.9	26.6	12.2	10.5
	Finetuned	29.1	36.5	26.9	12.5	10.2
	FT + Mem + LoRA †	30.3	37.1	27.9	13.3	10.7
MagiV2	Pretrained	65.3	76.6	63.8	50.4	45.7
	Finetuned	65.9	77.4	63.6	50.5	45.7

Appendix

Table 13 : Cross-corpus transfer to Manga109 , 27 held-out volumes (Appendix A.16 ). Pretrained is the released backbone with no POPCharacters training at all. The two rows under it are POPCharacters -trained checkpoints scored zero-shot here, the no-memory baseline and then the best memory-block configuration by P1 mAP (

\dagger

), so the first gap in each block is what BNNeck fine-tuning on another corpus is worth. P4 is identity Rank-1, with P2 given at Rank-1 beside it.

	Method	Char-ID Acc (%)	mAP	Rank-1
VLMs	Qwen2.5-VL-3B [ 5 ]	1.11	–	–
	InternVL3-14B [ 5 ]	0.00	–	–
	Ovis2-8B [ 5 ]	0.85	–	–
Re-ID	TransReID (pre)	–	34.5	46.0
	InstructReID (pre)	–	30.7	50.9
	ReID5o (pre)	–	36.2	60.8

Appendix

Table 14 : Re:Verse benchmark. Character identification on Re:Zero. The VLM rows carry the character-identification accuracy published with the benchmark [ 5 ] . Every Re-ID row is our own harness (Appendix A.16 ), the mean over five seed draws, so the pretrained rows and the trained rows are directly comparable. The memory block is worth

+0.1

mAP here and costs 1.6 Rank-1, on a single series.

Backbone	Rule	#clusters	Purity	Hung. Acc	ARI	NMI
TransReID	fixed $\tau_{\mathrm{nov}}=0.55$	95.8	60.2	15.1	2.0	21.8
	variance-adaptive	111.2	62.3	13.8	1.6	22.9
	density-aware	180.6	69.5	8.6	0.9	27.8
	cohesion-relative	144.0	65.9	11.4	1.4	25.8
	graph community detection	21.8	51.0	17.5	2.6	13.0
MagiV2	fixed $\tau_{\mathrm{nov}}=0.55$	38.9	69.6	46.6	24.0	34.1

Appendix

Table 15 : P3 decision rules on full per-series streams (finetuned TransReID and MagiV2 from the first training run, POPCharacters test series, macro over 8 series). Best ARI per backbone in bold . No alternative improves on the fixed rule at a comparable cluster count: on MagiV2 every rule is below it, and on TransReID graph community detection edges past it only by collapsing to a quarter of the clusters. These rules are reference instantiations of the decision, not baselines of the framework, and no claim in the paper depends on their ranking.

	Finetuned					Trained with the memory block
Backbone	#cl.	Pur.	Hung.	ARI	NMI	#cl.	Pur.	Hung.	ARI	NMI
TransReID	95.8	60.2	15.1	2.0	21.8	90.1	59.3	16.3	2.2	21.3
MagiV2	38.9	69.6	46.6	24.0	34.1	39.9	69.9	47.4	24.5	34.5
MagiV3	79.9	65.2	20.7	5.9	26.0	69.0	64.1	23.0	6.7	24.6
InstructReID	278.8	81.0	10.5	2.1	33.3	260.6	78.8	11.0	2.2	32.2
ReID5o	213.8	76.0	10.6	1.8	31.2	199.1	74.3	10.5	1.8	30.4

Appendix

Table 16 : P3 on full per-series streams, all five backbones , fixed rule at

\tau_{\mathrm{nov}}=0.55

, macro over the 8 held-out POPCharacters series against 8.8 identities per series, first training run. Every crop of every series is streamed in reading order. P3 has no gallery to initialise memory from, so the block is bypassed and the right-hand columns score the BNNeck trained alongside it.

		Seq-R (random seeding)					Seq-T (chronological seeding)
Backbone	Config	Static	Pred.	Oracle	Wrong-app.	Contam.	Static	Pred.	Oracle	Wrong-app.	Contam.
TransReID	Finetuned	17.4	15.0	41.7	85.0	84.2	12.8	12.6	42.2	87.4	83.3
	FT + Mem	17.2	15.7	41.4	84.3	82.8	13.2	13.6	41.8	86.4	84.0
MagiV2	Finetuned	36.8	36.3	59.1	63.7	67.8	36.6	37.9	59.0	62.1	69.8
	FT + Mem	37.2	35.9	59.8	64.1	68.7	35.3	36.9	60.4	63.1	68.3
MagiV3	Finetuned	22.7	19.9	49.9	80.1	80.0	17.4	17.7	50.2	82.3	82.0

Appendix

Table 17 : P4 under its three update policies at

k=1

and

B_{\max}=50

on the 8 held-out POPCharacters test series, identity Rank-1, mean over three training runs and five seed draws. Static is the unchanged P2 gallery, predicted adds each query under its top-1 match (the protocol’s rule) and oracle adds each query under its true character. Wrong-app. is the fraction of added entries that carry the wrong character and contam. the mislabelled fraction of the grown gallery at stream end, both under the predicted policy. Finetuned trains a BNNeck head on the frozen backbone, with no LoRA; FT + Mem adds the memory block to it. The three policies share one seed draw per run.

Figure 9 : P4 over the stream at

k=1

and

B_{\max}=50

with the memory block, mean over three training runs, five seed draws and the 8 test series. Left and centre: identity Rank-1 in each quarter of the stream, pooled over the five backbones, for the static gallery, growth by the model’s own top-1 (predicted) and growth by the true label (oracle). Right: the share of the grown gallery that is mislabelled at the end of the stream under predicted growth.

Configuration	0 (=P2)	5	10	25	50	100	unbounded
TransReID, chronological	13.2	13.2	14.2	13.0	13.6	14.6	14.9
TransReID, random	17.2	16.1	16.4	15.9	15.7	15.4	15.4
MagiV2, chronological	35.3	35.8	33.4	35.5	36.9	37.5	37.4
MagiV2, random	37.2	35.4	34.9	35.9	35.9	36.1	36.1

Appendix

Table 18 : Sensitivity of P4 to the buffer cap

B_{\max}

(identity Rank-1 at

k=1

on the 8 held-out POPCharacters series, mean over three training runs).

B_{\max}=0

is the static P2 gallery. No cap turns growth by predicted labels into a gain over that gallery: under random seeding every cap is below it, and under chronological seeding MagiV2 recovers about two points as the cap loosens without reaching the oracle’s range. The cap is therefore not the free parameter that decides whether growth pays. Acceptance is (Section 6 ).

		TransReID FT		TransReID +Mem		MagiV2 FT		MagiV2 +Mem
Box condition	IoU	P1	P2	P1	P2	P1	P2	P1	P2
clean box	1.00	37.4	38.6	37.5	38.4	51.2	54.4	51.7	54.5
shift 10%	0.83	+0.1	+0.2	+0.2	-0.1	-0.0	+0.4	+0.2	+0.0
shift 20%	0.69	-0.2	-0.3	-0.3	-0.4	-0.9	-1.4	-0.5	-1.4
shift 30%	0.58	-0.7	-0.4	-0.8	-0.2	-2.4	-2.7	-2.1	-2.7
tight $0.7\times$	0.49	+0.6	+1.5	+0.6	+1.6	-0.6	-0.2	-0.5	-0.7

Appendix

Table 19 : Bounding-box displacement. P1 and P2 Seq-R at

k{=}1

, three seed draws, macro over the 8 test series. The first row is absolute mAP and every later row is the change from it in the same column. IoU is the mean overlap of the displaced box with the true one.

	TransReID FT		TransReID +Mem		MagiV2 FT		MagiV2 +Mem
Pixel condition	P1	P2	P1	P2	P1	P2	P1	P2
clean pixels	37.4	38.6	37.5	38.4	51.2	54.4	51.7	54.5
jitter 10%	+0.0	+0.3	+0.1	-0.3	-0.1	+0.1	-0.0	+0.0
jitter 20%	-0.2	+0.1	-0.2	-0.5	-0.5	-0.3	-0.5	-0.4
blur $\sigma{=}2$	+0.2	+0.6	+0.1	+0.3	-1.2	-1.2	-1.2	-1.3
blur $\sigma{=}4$	+0.2	+0.4	-0.1	-0.7	-4.9	-5.4	-5.0	-6.3

Appendix

Table 20 : Pixel corruption (same setting as Table 19 ). Blur and occlusion, not box placement, are what separate the two backbones.

Backbone	Configuration	P1 mAP	P1 R-1	P2-R@1 mAP	$\Delta$ P1 mAP vs full
MagiV2	Finetuned (no memory)	51.20	57.17	54.44	$-0.54^{*}$
	Full memory block	51.75	56.29	54.58	–
	$-$ working memory	51.15	57.28	54.53	$-0.59^{*}$
	$-$ episodic memory	52.00	56.26	54.62	$+0.26$
	$-$ ID-drop	51.91	56.25	54.62	$+0.16$
	$-$ memory-consistency loss	51.92	56.60	54.64	$+0.18^{*}$

Appendix

Table 21 : Per-component ablation of the memory block on the two backbones where it earns anything, macro over the 8 held-out POPCharacters test series over three training runs. Each row removes one component from the full memory block; no row carries LoRA.

\Delta

is against the full memory block, paired over the 24 (series, training run) cells, and ∗ marks

p<0.05

under a two-sided exact sign test on those pairs. Working memory carries the whole effect: removing it returns P1 mAP to the finetuned baseline on both backbones, while removing episodic memory, ID-drop or the memory-consistency loss moves P1 mAP upward .

Figure 10 : P3 novelty-threshold sensitivity on full per-series streams: predicted clusters per series (log scale; dashed, the 8.8 true identities), Purity, NMI and ARI against

\tau_{\mathrm{nov}}

for the five finetuned backbones, first training run, macro over the 8 test series. The dotted line is the reference

\tau_{\mathrm{nov}}=0.55

. Purity and NMI rise as the stream fragments, ARI falls or peaks early, and MagiV2 leads on ARI at every threshold.

Dataset	Split	#Series	#Chars	#Crops	Avg C/Ch
POPCharacters	Train	13	198	7,668	38.7
	Development	2	10	873	87.3
	Test	8	70	4,058	58.0
Manga109	Test	27	784	29,315	37.4
Re:Verse	Test	1	12	1,825	152.1

Appendix

Table 22 : Dataset statistics. Splits are series-disjoint, and volume-disjoint on Manga109. Avg C/Ch is the average crops per character. No model is trained on Manga109 or Re:Verse: every number on them is zero-shot.

Figure 11 : Crops per series in POPCharacters , coloured by split: 13 training series, 2 development series and the 8 test series. Character counts above bars.

Figure 12 : Crop count distribution per series. Series with large casts (left) exhibit heavier tails in per-character crop counts.

Explore similar work

Character Identity is not Speaker Identity: KyaraBench and KyaraEmbed for Character Verification
Oct 5, 2026Joonyong Park, Jerry Li
Spellbrush, CA, USA
Consistency Has a Computable Blind Spot: A Commutation Theory of Label-Free Reliability for Vision-Language Figure Reading
Aug 6, 2026Rasul Khanbayov, Hasan KurbanExplanation Consistency Score Blind Spots
College of Science and Engineering, Hamad Bin Khalifa University, Doha, Qatar
End-to-End Text Line Detection and Ordering
Jun 2, 2026Benjamin KiesslingScene Text Recognition Text Analysis and Detection
ALMAnaCH · ALMAnaCH, Inria, France

Oct 5, 2026, eess.ASJ/K move · Enter open · S save

Character Identity is not Speaker Identity: KyaraBench and KyaraEmbed for Character Verification

Joonyong Park, Jerry Li

Spellbrush, CA, USA

Open paper AskPDF arXiv

A dubbed character keeps its identity while the voice actor changes, so character identity and speaker identity are distinct properties of one recording, yet speaker verification measures only the latter. To address this gap, we propose KyaraBench, a benchmark that scores character voice directly instead of speaker voice, built from 85 human-audited identities in a dubbed anime corpus. It poses two challenging conditions: one that swaps the performer under a fixed character, and one that fixes the performer under changing characters. Listening studies with 78 participants provide human reference scores for cross-performer verification and same-actor discrimination. Speaker-verification baselines show increased errors under these character-specific conditions. We then train KyaraEmbed, a compact encoder using multilingual character supervision, same-actor negatives, and a language-alignment term. The model achieves the best performance on all character-specific conditions in the main comparison. We release the benchmark, protocol, and encoder publicly.

Abstract

Explore similar work

Character Identity is not Speaker Identity: KyaraBench and KyaraEmbed for Character Verification

Consistency Has a Computable Blind Spot: A Commutation Theory of Label-Free Reliability for Vision-Language Figure Reading

End-to-End Text Line Detection and Ordering