EmoTrace-Bench: Diagnosing Emotional Expression in Text-to-Video Generation through Controlled Cues and Intensity

Yeheng Wang; Liudan Shi; Yi Zheng; Jingke Li; Shiyuan Liu; Xin Tao; Yuanxing Zhang; Guoyu Sun

cs.CVOct 11, 2026

EmoTrace-Bench: Diagnosing Emotional Expression in Text-to-Video Generation through Controlled Cues and Intensity

Authors: Yeheng Wang, Liudan Shi, Yi Zheng, Jingke Li, Shiyuan Liu, Xin Tao, Yuanxing Zhang, Guoyu Sun

Organizations: Communication University of China, Beijing, China · Kling Team, Beijing, China

Abstract

Text-to-video (T2V) models are increasingly used to generate videos with targeted emotional expressions, yet existing evaluation methods mainly rely on holistic video-text alignment and provide limited insight into whether failures arise from target-emotion expression, intensity control, or specific affective cues. We introduce EmoTrace-Bench, a fine-grained diagnostic benchmark that controls emotion intensity and selected affective cues in prompts and evaluates generated videos at multiple levels. EmoTrace-Bench contains 448 controlled prompts covering eight emotions, two intensity levels, and diverse cue conditions. Using 12 T2V models, we generate 5,376 videos and evaluate target-emotion presence, perceived intensity, and ten audiovisual dimensions of emotional expression with an automated evaluator calibrated and validated against human annotations on a subset of the data. Results reveal substantial differences across models and prompt designs. Prompts that explicitly describe a character's emotional reaction are followed more reliably than prompts that describe only events or scenes, whereas adding more affective cues or using more explicit cinematic audiovisual language does not consistently improve emotional generation. These results illustrate how aggregate alignment scores may overlook localized differences in target-emotion expression, intensity control, and cue realization. EmoTrace-Bench provides a traceable and interpretable framework for diagnosing emotional video generation without modifying existing generation pipelines, and we release the prompts and the automated evaluation rules.

Figures & tables

Figure 1: EmoTrace-Bench at a glance. Left: the three-level evaluation taxonomy—whether the target emotion is expressed, whether its perceived intensity is weak or strong, and which of ten expression channels (character, visual form, audio, context, temporal) provide the evidence. Top right: correct content does not guarantee emotional expression; a model must infer the emotion from context and reaction and realize the requested intensity, and existing benchmarks do not diagnose where this fails. Bottom right: the four capability facets examined in Section 4—breadth across emotions and prompt conditions, depth of intensity control and cue support, and explicit versus implicit cue expression.

Affective target

Controlled input

Diagnostic output

Human validation

Benchmark

EmoVid ( Qiu et al., 2026 )

\bullet

\circ

\odot

\circ

\circ

\odot

\circ

\bullet

\bullet

\circ

\odot

\bullet

HVEval ( Wu et al., 2025 )

\odot

\circ

\odot

\circ

\circ

\odot

\circ

\odot

\bullet

\circ

\odot

\bullet

HUG-VIS ( Ma et al., 2026 )

\bullet

\circ

\circ

\circ

\circ

\odot

\circ

\bullet

\bullet

\circ

\odot

\bullet

Stable Cinemetrics ( Chatterjee et al., 2025 )

\circ

\circ

\odot

\odot

\circ

\bullet

\odot

\circ

\bullet

\circ

\odot

\odot

FilmBench ( Wang et al., 2026b )

\odot

\circ

\odot

\circ

\circ

\odot

\circ

\odot

\bullet

\circ

\odot

\bullet

Table 1: Capability coverage of representative video-generation benchmarks. Filled, dotted, and open circles denote full, partial or domain-restricted, and absent support, respectively.

Figure 2: Overview of the EmoTrace-Bench pipeline. Prompts from the controlled suite (top) are sent unchanged to every model; the middle row shows matched weak/strong generations for one fear prompt pair across models. All videos are scored by a single automated evaluator that is frozen after calibration against human labels (left) and applied through the two-stage affective evaluation (bottom).

Figure 3: Emotion generation across 12 T2V models. (a) Overall performance: target-emotion accuracy (Emotion), joint emotion-and-intensity accuracy (Emotion + intensity; E+I in the text), Visual 9D and Full 10D cue support, and bootstrap group; models within a group are not ordered. (b) Target-emotion accuracy for each of the eight emotions. Scores are obtained from one blinded automated evaluation per video.

Figure 4: Overall (left) and implicit (right) cue-expression profiles of 12 T2V models. Values are relative to the best model in each channel.

Reference	Target emotion	Conditional intensity	Joint three-way	Ten dimensions
Development, first-pass human reference	88.1 / .67	79.1 / .58	73.4 / .59	83.0 / .63
Test, first-pass human reference	84.7 / .59	88.3 / .76	77.0 / .65	80.5 / .59
Development, reviewed human reference	89.9 / .72	93.0 / .86	84.8 / .76	85.1 / .67
Test, reviewed human reference	90.2 / .76	92.0 / .84	87.1 / .81	83.0 / .64

Table 2: Agreement between the automated evaluator and the human reference.

Appendix figures & tables23 assets

Supplementary material from the paper’s appendix.

Appendix

Public model designation	Access	Videos	Native audio
Seedance-2.5	hosted	448	yes
Veo-3.1	hosted	448	yes
Wan-3.0	hosted	448	yes
HappyHorse-1.1	hosted	448	yes
MiniMax-H3-Max (fal)	hosted	448	yes
Kling-3.0 †	hosted	448	yes

Appendix

Table A.1: Systems included in the 5,376-video evaluation. Each model generates one video for each of the 448 prompts. “Access” distinguishes systems reached through a hosted interface from open-weight systems run locally; “Native audio” indicates whether the generated output contains a usable model-produced audio track.

Field	Protocol
Prompt submission	Identical frozen English prompt for all 12 models
Samples per prompt	One retained generation per model–prompt pair; no best-of- $n$ selection
Duration	Approximately five seconds (provider-native duration closest to the target), continuous single-shot target; the full clip is evaluated and duration does not enter any score
Resolution	Provider-recommended native output; not forcibly matched across models
Frame rate	Provider-native output; automated evaluation samples at 8 fps
Audio	Original model-produced track retained; sound excluded when no usable native track exists

Appendix

Table A.2: Generation controls and cross-model treatment. “Native” means that a property is preserved from the provider output rather than normalized by post-processing.

Metric	Seed A	Seed B	Seed C	Range	SD
Target emotion (%)	75.45	72.54	76.12	3.57	1.55
Emotion + target intensity (%)	47.77	45.54	49.33	3.79	1.56
Visual 9D macro (%)	72.90	72.92	72.74	0.18	0.08
Full 10D macro (%)	69.72	70.03	69.72	0.31	0.15

Appendix

Table A.3: Kling-3.0 score variation across seed settings. All batches use the same 448 prompts and exposed generation configuration; only the generation seed is changed. Visual 9D and Full 10D are emotion-conditioned macro averages on each batch’s own emotion-matched subset. SD is the descriptive standard deviation over the three settings (divisor 3) and does not estimate variance over unseen seeds.

Comparison	Metric	$\Delta$	95% CI	McNemar $p$
A $\rightarrow$ B	Target emotion	$-2.90$	$[-6.92,+1.12]$	.203
	Emotion + target intensity	$-2.23$	$[-6.92,+2.23]$	.373
	Visual 9D macro	$+0.02$	$[-2.39,+2.53]$	–
	Full 10D macro	$+0.31$	$[-2.11,+2.68]$	–
A $\rightarrow$ C	Target emotion	$+0.67$	$[-3.12,+4.46]$	.818
	Emotion + target intensity	$+1.56$	$[-2.90,+6.25]$	.558

Appendix

Table A.4: Prompt-matched pairwise differences between seed batches.

\Delta

is the right-hand seed minus the left-hand seed in percentage points; intervals are 95% percentile intervals from 5,000 paired prompt bootstrap replicates;

p

is from the two-sided exact McNemar test, applicable to the two binary metrics only.

Comparison	Target emotion	Three-level intensity	Ten dimensions
A vs. B	80.1 / .484	70.1 / .527	81.9 / .615
A vs. C	83.3 / .544	70.5 / .532	82.1 / .619
B vs. C	83.9 / .580	72.1 / .560	82.8 / .634

Appendix

Table A.5: Same-prompt agreement between seed batches. Values are agreement (%) / Cohen’s

\kappa

between different generated videos for the same prompt, not repeated evaluations of one video.

Template	Family	Complexity	Segments written	$\|\mathrm{Exp}\|$	$\|\mathrm{atoms}\|$	Words	Explicit channels	Implicit channels
EMO	emotion-label baseline	atomic	emotion label	0	1	15	–	C1, C2, V1–V4, X1, X2, T1
FACE	single channel	atomic	face	1	1	26.5	C1	C2, V1–V4, T1
BODY	single channel	atomic	body	1	1	25.5	C2	C1, V1–V4, T1
SCENE	single channel	atomic	scene appraisal	1	1	24.5	X1	C1, C2, V1–V4, T1
EVENT	single channel	atomic	event appraisal	1	1	24.5	X2	C1, C2, V1–V4, T1
FACE_BODY	paired addition	composed	face + body	2	2	42.5	C1, C2	V1–V4, T1

Appendix

Table B.1: The 14 prompt templates. Every template contributes 32 prompts: eight emotions, two scenes, and two intensity levels. “Words” is the median prompt length. Visual-form segments jointly specify V1–V4; these four channels are not independently manipulated. A dash means that no channel is explicit.

Template	C1	C2	V1	V2	V3	V4	X1	X2	T1
EMO	I	I	I	I	I	I	I	I	I
FACE	E	I	I	I	I	I	–	–	I
BODY	I	E	I	I	I	I	–	–	I
SCENE	I	I	I	I	I	I	E	–	I
EVENT	I	I	I	I	I	I	–	E	I
FACE_BODY	E	E	I	I	I	I	–	–	I

Appendix

Table B.2: Prompt-side cue roles for all templates. E = explicitly written; I = implicit (not written but observed); – = outside the intended scope of that template. A1 is omitted because sound is scoring-only.

	Pair	Operation	Intended change	Tested
1	SCENE $\rightarrow$ SCENE_REACT	add	add facial and bodily reaction to the same scene appraisal	yes
2	EVENT $\rightarrow$ EVENT_REACT	add	add facial and bodily reaction to the same event appraisal	yes
3	FACE $\rightarrow$ FACE_BODY	add	add a body segment to the same facial description	yes
4	BODY $\rightarrow$ FACE_BODY	add	add a facial segment to the same body description	yes
5	CONTEXT_MOOD $\rightarrow$ CONTEXT_HARD	replace	experiential atmosphere $\rightarrow$ explicit cinematic form, character present	yes
6	EMPTY_MOOD $\rightarrow$ EMPTY_HARD	replace	experiential atmosphere $\rightarrow$ explicit cinematic form, empty scene	yes

Appendix

Table B.3: Seven controlled template pairs. Text not named in the “Intended change” column is preserved within the pair. “Tested” indicates membership in the pre-specified McNemar/Holm family.

Figure B.1: Six prompt conditions for one fear scene, shown as strong–weak pairs with the intensity-bearing span highlighted. Rows 1–6 correspond to the EMO, FACE, SCENE, SCENE_REACT, CONTEXT_MOOD, and CONTEXT_HARD templates (Table B.1 ); rows 3–4 and rows 5–6 are contrasts 1 and 5 in Table B.3 .

Figure B.2: Frame sequences from three models for the same strong-intensity prompt within each of three emotions. Although every video depicts the requested scene and character, the realized expression ranges from clearly strong to nearly neutral, so emotion presence and intensity are scored separately.

Check	Result
Total structured prompt records	448
Strong / weak members	224 / 224
Unique prompt texts	448 / 448
Unique SHA-256 prompt hashes	448 / 448
Intensity pairs passing the pairing contract	224 / 224
Literal reuse of shared segments	372 / 384

Appendix

Table B.4: Structural audit of the 448-prompt set. Structural checks verify the frozen construction contract.

Group	ID	Channel	Direct object of judgment	Channel-specific prerequisite
Character performance	C1	Facial expression	Joint configuration of eyes, brows, mouth, cheeks, and jaw	An identifiable, readable face
	C2	Bodily behavior	Posture, gesture, movement direction and energy, or meaningful stillness	An identifiable, readable body
Visual form	V1	Lighting	Brightness, contrast, shadow, direction, hardness, and their change over time	Visual content is readable
	V2	Color	Palette, warm–cool balance, saturation, color contrast, and change	Chromatic information is present
	V3	Composition	Placement, negative space, visual weight, balance, occlusion, and depth	Spatial configuration is visible
	V4	Visual dynamics	Camera motion, recomposition, stability, cuts, and meaningful holds	Sufficient temporal extent

Appendix

Table C.1: The ten affective cue channels, grouped by source, with their operational definitions. Each channel receives a binary judgment. A positive label requires observable evidence that specifically supports the target emotion, subject to the shared and channel-specific rules below.

Metric	Denominator	Treatment of missing or inapplicable values
Emotion	all valid videos ( $448$ per model)	exclude and separately count unscorable process failures
E+I	all valid videos ( $448$ per model)	emotion failure receives no joint credit; intensity is not forced
Cue channel $d$	videos with $M=1$	exclude a channel only when no valid object exists, such as A1 without native audio
Visual 9D	mean of nine visual/context/temporal channel rates	always reported for a valid model run
Full 10D	mean of ten channel rates	not reported for silent models
Paired control	all 224 strong–weak pairs	a pair with either emotion absent is a failed pair, not an intensity reversal

Appendix

Table C.2: Exact metric denominators.

Reference	Target emotion	Intensity	Ten dimensions
Development: pairwise agreement (%)	85.7	77.9	80.8
Development: Fleiss’ $\kappa$	.65	.50	.60
Reviewed reference vs. trained majority (%)	90.4	90.9	94.2
Test: two-annotator agreement (%)	81.7	75.5	80.2
Test: Cohen’s $\kappa$	.54	.50	.58

Appendix

Table D.1: Agreement among human judgments. Development-set values summarize five trained annotators. Test-set agreement is computed only between the two full-coverage annotators; the third, senior annotator serves only as a reviewer.

Item	Development	Test
First-pass agreed fields	8,899	1,780
First-pass disputed fields	317	524
Tie-broken by trained annotators	206	–
Excluded from first-pass reference (no consensus)	111	524
Fields changed at second review	308	297
Second-review annotators	same two senior annotators	a third, senior annotator

Appendix

Table D.2: First-pass consensus and second-review statistics. Counts refer to fields.

Reference	Target emotion	Conditional intensity	Joint three-way	Ten dimensions
Development, first-pass human reference	$88.1{\pm}0.7$ / $.670{\pm}.021$	$79.1{\pm}0.1$ / $.578{\pm}.003$ (637)	$73.4{\pm}0.6$ / $.588{\pm}.009$	$83.0{\pm}0.2$ / $.625{\pm}.003$ (7,424)
Test, first-pass human reference	84.7 / .589 (183)	88.3 / .764 (94)	77.0 / .652	80.5 / .585 (1,489)
Development, reviewed human reference	$89.9{\pm}0.9$ / $.716{\pm}.024$	$93.0{\pm}1.3$ / $.858{\pm}.026$ (641)	$84.8{\pm}1.8$ / $.764{\pm}.028$	$85.1{\pm}0.3$ / $.669{\pm}.007$ (7,424)
Test, reviewed human reference	90.2 / .762 (224)	92.0 / .840 (125)	87.1 / .806	83.0 / .643 (1,639)

Appendix

Table D.3: Complete agreement results between the automated evaluator and the human reference. Values are agreement (%) / Cohen’s

\kappa

; development-set values are mean

\pm

standard deviation over three independent runs (the 12-model run plus two stability runs); test-set values are from the single run. Parentheses give valid counts.

Dimension	Development	Test
Facial expression	87.9 / .74 / 768 / 64 $\rightarrow$ 63	84.6 / .69 / 156 / 58 $\rightarrow$ 52
Bodily behavior	85.1 / .70 / 768 / 53 $\rightarrow$ 59	84.1 / .68 / 157 / 51 $\rightarrow$ 52
Lighting	82.4 / .55 / 896 / 74 $\rightarrow$ 72	77.7 / .50 / 224 / 70 $\rightarrow$ 63
Color	85.3 / .64 / 896 / 71 $\rightarrow$ 71	83.5 / .64 / 224 / 67 $\rightarrow$ 63
Composition	84.9 / .55 / 896 / 80 $\rightarrow$ 77	79.3 / .30 / 174 / 93 $\rightarrow$ 74
Visual dynamics	84.0 / .64 / 896 / 64 $\rightarrow$ 69	86.0 / .72 / 186 / 49 $\rightarrow$ 48

Appendix

Table D.4: Per-dimension agreement against the reviewed human reference. Each cell reports agreement (%), Cohen’s

\kappa

, valid count, and the human-to-automated pass rates; run averaging follows Table D.3 .

Reference / task	Qwen3.8-Omni-Flash	Gemini 3.1 Pro
First-pass: target emotion	81.4 / .509	83.6 / .574
First-pass: joint three-way	70.3 / .542	73.6 / .602
First-pass: conditional intensity	84.3 / .679	82.8 / .653
First-pass: ten dimensions	83.1 / .627	79.9 / .575
Reviewed: target emotion	84.4 / .627	88.4 / .721
Reviewed: joint three-way	74.2 / .610	80.4 / .705

Appendix

Table D.5: Test-set agreement of two evaluator backbones with the human reference. Values are agreement (%) / Cohen’s

\kappa

; both evaluators run at 5 fps under the frozen protocol.

Pair	Emotion	Joint three-way	Ten dimensions	Visual 9D	Sound
Qwen3.8-Omni-Flash vs. Gemini (5 fps)	.640	.622	.678	.676	.678
Qwen3.8-Omni-Flash vs. Gemini (8 fps)	.679	.630	.680	.682	.660
Gemini 8 fps vs. 5 fps	.875	.844	.861	–	.893

Appendix

Table D.6: Direct agreement between evaluator runs on the same test-set videos. Values are Cohen’s

\kappa

. The last row compares the 8 fps setting used in the main results with the 5 fps control.

		Character		Visual form				Audio	Context		Temporal
Model	$N$	C1	C2	V1	V2	V3	V4	A1	X1	X2	T1
Seedance-2.5	390	86.4	70.1	85.1	84.9	89.7	93.6	82.6	89.8	58.6	93.2
Veo-3.1	380	83.6	74.1	76.8	79.0	87.4	44.5	61.0	91.7	62.3	91.4
Wan-3.0	358	82.2	68.1	76.5	79.6	85.8	33.8	78.5	91.6	46.4	87.8
HappyHorse-1.1	357	80.7	73.1	72.8	76.2	82.6	86.8	75.3	88.8	51.9	85.9
MiniMax-H3-Max (fal)	333	81.4	66.4	74.2	80.8	85.0	33.0	52.2	89.0	49.0	88.9

Appendix

Table E.1: Cue-level expression profiles of 12 T2V models. Each score is the percentage of target-emotion-successful videos in which the corresponding channel supports the target emotion.

N

is the number of such videos for each model.

	Character		Visual form				Audio	Context		Temporal
Model	C1	C2	V1	V2	V3	V4	A1 ¶	X1 §	X2 §	T1 ¶	Implicit 9D
Seedance-2.5	71.5	54.9	81.2	81.6	88.1	94.6	82.6	90.3	29.0	93.2	76.1
Veo-3.1	67.4	57.6	71.0	73.0	85.3	36.9	61.0	93.5	35.5	91.4	68.0
Wan-3.0	63.5	48.5	70.3	74.2	81.8	25.0	78.5	90.0	6.7	87.8	60.9
HappyHorse-1.1	70.2	61.1	66.8	70.2	79.0	86.5	75.3	83.3	36.7	85.9	71.1
MiniMax-H3-Max (fal)	61.3	42.5	68.3	74.9	81.5	23.8	52.2	78.6	17.9	88.9	59.7

Appendix

Table E.2: Implicit cue-expression profiles of 12 T2V models. Each score is the percentage of target-emotion-successful videos, among those whose prompt does not explicitly specify the channel, in which that channel supports the target emotion. Implicit 9D is the mean over the nine visual, contextual, and temporal channels.

Model	Correctly ordered	Collapsed	Reversed	Not comparable
Seedance-2.5	27.7	47.3	5.8	19.2
Veo-3.1	28.6	48.2	1.3	21.9
Wan-3.0	26.8	43.8	0.9	28.6
HappyHorse-1.1	29.5	36.6	3.1	30.8
MiniMax-H3-Max (fal)	23.2	36.6	1.3	38.8
Kling-3.0	19.6	39.7	1.3	39.3

Appendix

Table E.3: Outcomes of the 224 strong–weak prompt pairs per model. All values are percentages of pairs and sum to 100 within a row.

Emotion-label baseline
Template	Seedance-2.5	Veo-3.1	Wan-3.0	HappyHorse-1.1	MiniMax-H3-Max (fal)	Kling-3.0	MiniMax-H3	LTX-2.5	Wan-2.2	HunyuanVideo-1.5	Kandinsky-5.0-T2V-Pro	LongCat-Video-13.6B	Mean	Sig.
EMO	66	72	81	75	72	62	62	69	62	50	44	25	62
Character performance
FACE	69	81	50	62	69	66	56	53	47	53	34	41	57
BODY	56	41	47	53	38	47	25	50	28	28	38	25	40
FACE_BODY	69	69	66	81	62	56	66	78	50	62	66	41	64

Appendix

Table E.4: Joint emotion-and-intensity accuracy (%) by prompt template and model. Regular rows report E+I over the 32 videos of each template. Italic

\Delta

rows report the paired difference (in percentage points) of the row immediately above from the named template, for the six pre-specified contrasts; ∗ denotes

p<0.05

under an exact paired McNemar test after model-wise Holm correction over the six contrasts. “Mean” is the 12-model mean; “Sig.” counts models with a significant difference.

T2V Generation

Text-to-video generation creates moving visual scenes from natural-language prompts, with growing control over motion, identity, events, and style.

See all 111 papers

VTR-Bench: A Systematic Benchmark for Evaluating Visual Text Rendering in Video Generation
Oct 1, 2026Yu Huang, Jungang Li, Zhiyuan Wang +8Agentic Video Generation Video Generation
Department of Data Science, City University of Hong Kong · The Hong Kong University of Science and Technology (Guangzhou) · Westlake University +3
EmoGenie: Benchmarking and Evaluating Expressed and Evoked Emotion in T2I Generation with MLLMs
Oct 13, 2026Zixuan Xing, Lancheng Gao, Shengyan Li +4T2I Generation Evaluation Multimodal Emotion Recognition
Shanghai Jiao Tong University
TextVBench: Evaluating Text Rendering and Preservation in Video Generation
Oct 9, 2026Ruiqiang Zhang, Hengyi Wang, Chang Liu +3Video Generation Evaluation T2V Generation
Anhui Province Key Laboratory of Digital Security, University of Science and Technology of China

Oct 1, 2026J/K move · Enter open · S save

VTR-Bench: A Systematic Benchmark for Evaluating Visual Text Rendering in Video Generation

Yu Huang, Jungang Li, Zhiyuan Wang, Yonghua Hei+7

Department of Data Science, City University of Hong Kong · The Hong Kong University of Science and Technology (Guangzhou) · Westlake University+3

Agentic Video Generation Video Generation Video Generation Evaluation

Open paper AskPDF arXiv

Recent video generation models can produce highly realistic videos from natural language instructions, with visual quality approaching cinematic standards. Existing evaluation benchmarks, however, predominantly assess visual quality, aesthetic appeal and physical plausibility, while paying limited attention to text, an essential medium for conveying information in everyday scenes. A generated video may appear visually compelling and feature lifelike subjects, yet still render the text within the scene incorrectly. To address this overlooked dimension, we introduce VTR-Bench, a systematic benchmark for evaluating the Visual Text Rendering capabilities of video generation models. VTR-Bench situates text within concrete application scenarios, such as advertisements and scientific videos, with 300 carefully constructed prompts spanning five scenario categories. We develop an automated evaluation pipeline with human alignments that separately assesses text fidelity through carrier-specific transcription and scene and motion requirements through a prompt-specific chain of query. Beyond evaluation, we introduce a Keyframe-Guided Agentic Framework in which a Director agent coordinates image and video generation with visual evaluation, guiding iterative refinement and candidate selection through visual feedback. Experiments on 11 state-of-the-art models reveal widespread difficulties in accurately rendering scene text, with the best-performing model recording an overall word error rate (WER) of 0.250. We further analyze text rendering failures to characterize the challenges faced by current video generation models. These findings highlight visual text rendering as a key challenge for video generation and demonstrate a practical path toward improvement. Code is available at https://github.com/hardenyu21/VTR-Bench.

Abstract

T2V Generation

Explore similar work

VTR-Bench: A Systematic Benchmark for Evaluating Visual Text Rendering in Video Generation

EmoGenie: Benchmarking and Evaluating Expressed and Evoked Emotion in T2I Generation with MLLMs

TextVBench: Evaluating Text Rendering and Preservation in Video Generation