A Large-scale Evaluation of Text-guided Models for Facial Editing

Rahul Nair; Saurav Pandit; Hannah Kerner

cs.CVAug 28, 2026

A Large-scale Evaluation of Text-guided Models for Facial Editing

Authors: Rahul Nair, Saurav Pandit, Hannah Kerner

Organizations: Arizona State University Tempe, AZ, USA · Prelight Inc New York, USA

Abstract

Facial appearance editing powers popular applications like FaceApp and Photoshop. Generative Adversarial Networks (GANs) and 3D Morphable Models (3DMMs) have been widely used for facial editing. GANs can perform varied facial edits (e.g., changing hair color, hairstyle), but often produce unstable edits. 3DMMs produce stable edits, but can only alter pose and facial expression. Recently, text-guided diffusion models like Nano Banana have become popular for image editing. Text-guided models are a compelling alternative to GANs and 3DMMs since they can produce both stable and varied image edits. While text-guided models have been widely tested for whole-scene edits (e.g., ``make the woman play a guitar''), they have not been comprehensively tested for facial editing. We conducted the first large-scale evaluation ( $\sim1$ M images evaluated) of six popular text-guided models on a sequential facial editing task. We present Face-Edit-Attributes, the largest collection of $169$ facial editing attributes focused on hair, accessories, and pose edits. We compared model performance using two popular celebrity face datasets: CelebA and CelebSET. Our results show that most models performed hair and accessory edits well, but struggled with editing pose. All models over-edit (e.g., changing hair color when asked only to change the hairstyle). We also evaluated demographic biases in each model. Our results show surprising biases in overediting: almost all models created more overedits for dark-skinned male faces and old faces. The code and data for our results (including our repository of $\sim 1$ M images) can be accessed blue{here}.

Figures & tables

Figure 1. An example editing session from our study. Each image goes through a fixed sequence of $4$ edits.

Figure 2. Face-Edit-Attributes: Leaf nodes (e.g., crimson red hair) are used to create the editing instructions.

Figure 3. Visually comparing good and bad scenarios for our evaluation metrics. Left: the top row (Nano-Banana) shows good identity preservation, while the bottom row (Flux-Dev) shows poor identity preservation. Center: the top row (Nano Banana) shows a correctly executed edit, while the bottom row (Flux-Dev) shows a poorly executed edit. Right: the top row (Nano Banana) shows a case for no overedits, while the bottom row (Qwen-Edit) indicates an overedit.

Model	Hair $\downarrow$	Acc. $\downarrow$	Pose $\downarrow$	Multi $\downarrow$	Avg. $\downarrow$
Kontext-Pro	24.60 0.16	25.37 0.14	37.48 0.28	32.57 0.17	28.68 0.10
Nano-Banana	23.08 0.12	33.51 0.14	32.43 0.15	32.06 0.16	29.93 0.08
Bagel-Edit	18.79 0.09	27.37 0.10	52.98 0.28	33.51 0.20	30.33 0.11
Qwen-Edit	28.04 0.20	27.90 0.14	47.59 0.22	34.82 0.17	32.73 0.10
Flux-Dev	89.13 0.06	89.58 0.06	89.68 0.09	89.40 0.06	89.41 0.04

Table 1.

\operatorname{Drop}_{\mathrm{FSS}}

(lower is better) across all test sets on CelebA and CelebSET. For each model, blue number indicates its best performance, while red indicates its worst performance across the four test sets. The Avg. column shows the average performance of a model across test sets. The bolded number indicates the overall best-performing model. Subscripts denote

95\%

confidence intervals.

Model	Hair $\uparrow$	Acc. $\uparrow$	Pose $\uparrow$	Multi $\uparrow$	Avg. $\uparrow$
Kontext-Pro	76.10 0.50	73.40 0.60	47.40 0.90	64.00 0.50	68.40 0.30
Nano Banana	87.90 0.30	84.70 0.40	62.60 0.80	75.10 0.50	79.60 0.30
Bagel-Edit	80.30 0.40	87.80 0.30	59.60 0.70	68.90 0.50	76.20 0.30
Qwen-Edit	76.60 0.50	79.90 0.40	67.50 0.80	64.70 0.50	72.90 0.30
Flux-Dev	30.30 0.50	33.70 0.50	31.40 0.80	30.20 0.50	31.40 0.30

Table 2.

SC

(higher is better) across all test sets on CelebA and CelebSET. For each model, blue number indicates its best performance, while red number indicates its worst performance across the four test sets. In the Avg. column, the bolded number indicates the best-performing model. Values scaled by

100

Intended Edit $\rightarrow$ Overedit	Models
bob cut $\rightarrow$ straight hair	BE, K-Pro, QE
sideburn mutton chops $\rightarrow$ handlebar mustache	BE, K-Pro, QE
bob cut $\rightarrow$ brown hair	F-Dev, QE
sideburn mutton chops $\rightarrow$ pencil mustache	K-Pro, QE
straight hair $\rightarrow$ brown hair	BE, QE

Table 3. Frequent overediting patterns across models. Nano-Banana (NB), SeedEdit (SE), Bagel-Edit (BE), Qwen-Edit (QE), Flux-Dev (F-Dev), Kontext-Pro (K-Pro).

Model	Hair $\downarrow$	Acc. $\downarrow$	Pose $\downarrow$	Multi $\downarrow$	Avg. $\downarrow$
Kontext-Pro	19.54 0.67	11.38 0.64	16.47 1.10	22.83 0.77	17.75 0.39
Nano Banana	15.02 0.59	10.42 0.71	10.18 0.83	17.62 0.69	13.79 0.35
Bagel-Edit	18.47 0.64	8.48 0.56	13.82 0.95	21.03 0.73	15.68 0.35
Qwen-Edit	24.87 0.76	13.66 0.74	19.87 1.05	27.25 0.90	21.63 0.43
Flux-Dev	66.42 1.33	68.38 1.31	47.76 1.58	65.31 1.29	64.00 0.70

Table 4. OER (lower is better) across all test sets on CelebA and CelebSet. For each model, blue number indicates its best performance, while red number indicates its worst performance across the test sets. The bolded number indicates the best-performing model. Values scaled by

100

	Accessories $\downarrow$		Hair $\downarrow$		Pose $\downarrow$		Multi-axis $\downarrow$
Model	Mean	$WD_{g}$	Mean	$WD_{g}$	Mean	$WD_{g}$	Mean	$WD_{g}$
Flux-Dev	F:89.7 M:89.4	0.4	M:89.2 F:89.1	0.4	F:90.0 M:89.2	0.8	F:89.5 M:89.3	0.5
Kontext-Pro	M:25.8 F:25.0	1.6	F:24.7 M:24.4	1.2	F:39.1 M:35.5	3.4	F:33.6 M:31.2	2.4
Nano Banana	M:34.5 F:32.8	1.8	M:24.3 F:22.1	2.3	M:33.0 F:32.0	1.0	M:33.0 F:31.3	2.2
Bagel-Edit	M:28.1 F:26.8	1.4	M:19.5 F:18.3	1.2	F:55.2 M:50.2	5.0	F:33.8 M:33.1	1.4
Qwen-Edit	M:29.1 F:27.0	2.5	F:28.8 M:27.1	2.0	F:50.8 M:43.6	7.3	F:35.3 M:34.2	1.8

Table 5.

\operatorname{Drop}_{\mathrm{FSS}}

% (lower is better) disparity on CelebA. For each model and test set, we reported mean

\operatorname{Drop}_{\mathrm{FSS}}

for Male (M) and Female (F). In each mean column, the left value is the mean of the worst performing group, and the right value is the mean of the best group.

WD_{g}

denotes the Wasserstein distance between the male and female

\operatorname{Drop}_{\mathrm{FSS}}

distributions. A superscript

*

marks non-significant values (

p>0.05

). WD > 3 is highlighted as it indicates gender bias.

	Accessories $\downarrow$			Hair $\downarrow$			Pose $\downarrow$			Multi-axis $\downarrow$
Model	Mean: 2 worst groups	$WD_{g}$	$WD_{s}$	Mean: 2 worst groups	$WD_{g}$	$WD_{s}$	Mean: 2 worst groups	$WD_{g}$	$WD_{s}$	Mean: 2 worst groups	$WD_{g}$	$WD_{s}$
Flux-Dev	FD:89.6 MD:85.5	0.9	3.7	FD:88.7 MD:85.8	0.5	4.2	FD:89.9 MD:85.2	1.3	4.0	FD:89.3 MD:85.1	0.8	3.6
Kontext-Pro	ML:28.1 FL:26.3	2.2	2.9	FL:22.0 ML:21.7	1.7	1.3	FD:33.5 FL:32.6	5.6	1.1 ∗	FL:31.0 ML:29.0	2.0	2.0
Nano-Banana	ML:27.2 MD:26.0	2.2	1.6	MD:17.5 FL:16.8	0.4 ∗	1.1	MD:25.1 FD:23.8	0.7 ∗	2.2	FL:24.1 ML:24.0	1.6	1.2
Bagel-Edit	ML:29.2 MD:27.2	2.4	1.9	MD:19.4 FD:18.8	1.9	1.6	MD:58.0 FD:57.6	1.2 ∗	2.2	MD:34.8 ML:33.8	1.3	0.9 ∗
Qwen-Edit	FL:50.6 FD:47.3	6.4	4.8	FL:52.0 FD:47.8	7.6	3.1	FD:45.4 FL:43.8	7.7	1.5	FL:49.8 FD:45.3	7.4	2.9

Table 6.

\operatorname{Drop}_{\mathrm{FSS}}

% (lower is better) disparity on CelebSET. We reported the mean

\operatorname{Drop}_{\mathrm{FSS}}

for the two worst intersectional groups. In each mean column, left value: worst; right value: second-worst mean group.

WD_{g}

and

WD_{s}

denote Wasserstein distances between gender and skin-tone groups, respectively. A superscript

*

marks non-significant values (

p>0.05

). ML: Male-Light, MD: Male-Dark, FL: Female-Light, FD: Female-Dark. WD

>3

is highlighted.

	Accessories $\uparrow$		Hair $\uparrow$		Pose $\uparrow$		Multi-axis $\uparrow$
Model	Mean	$WD_{g}$	Mean	$WD_{g}$	Mean	$WD_{g}$	Mean	$WD_{g}$
Flux-Dev	M:33.2 F:34.1	0.9 ∗	F:28.2 M:32.9	4.6	F:30.1 M:33.0	3.0	F:29.0 M:31.8	2.9
Kontext-Pro	F:73.4 M:73.5	1.8	M:75.5 F:76.7	2.0	F:46.5 M:48.6	2.6	F:63.5 M:64.6	1.2
Nano Banana	M:83.7 F:85.6	1.9	M:87.0 F:88.6	1.7	F:61.9 M:63.4	1.8	M:74.0 F:75.9	2.0
Bagel-Edit	M:87.5 F:88.0	0.6	M:79.0 F:81.2	2.3	M:58.8 F:60.2	1.6	M:68.8 F:69.0	1.4
Qwen-Edit	M:78.6 F:80.9	2.3	F:76.4 M:76.9	1.1	M:66.1 F:68.7	2.6	M:63.4 F:65.8	2.5

Table 7.

SC

(higher is better) disparity on CelebA. We reported mean

SC

values for Male (M) and Female (F). In each mean column, left: worst, right: best group.

WD_{g}

denotes the Wasserstein distance between the male and female

SC

distributions. A superscript

*

marks non-significant values (

p>0.05

). WD > 3 is highlighted. Values are scaled by

100

Accessories

\uparrow

Hair

\uparrow

Pose

\uparrow

Multi-axis

\uparrow

Model

Mean: 2 worst groups

WD_{g}

WD_{s}

Mean: 2 worst groups

WD_{g}

WD_{s}

Mean: 2 worst groups

WD_{g}

WD_{s}

Mean: 2 worst groups

WD_{g}

WD_{s}

Flux-Dev

ML:28.1 FL:28.9

0.8 ∗

1.0 ∗

FL:25.1 FD:26.6

5.5

1.3

FD:26.2 FL:27.8

2.9

1.2

FL:25.3 FD:27.3

4.1

0.3 ∗

Kontext-Pro

FL:72.2 ML:73.5

2.8

0.8 ∗

FL:74.6 ML:75.2

0.6 ∗

2.6

MD:35.1 FD:35.4

3.3

2.5

FD:60.4 FL:60.5

1.2 ∗

1.4

Nano Banana

FD:74.8 MD:80.9

1.5

4.4

ML:84.2 FL:86.8

2.4

FD:58.2 FL:58.8

2.1

1.0 ∗

MD:71.6 FL:73.1

1.1 ∗

Bagel-Edit

FL:85.7 FD:86.9

1.5

0.7 ∗

MD:76.7 FL:77.8

1.3

0.5 ∗

FL:59.6 FD:60.8

1.1 ∗

0.8 ∗

ML:67.2 FL:67.5

0.7 ∗

1.4

Qwen-Edit

ML:82.0 FL:82.4

1.2

FL:76.2 ML:79.7

2.9

2.7

ML:60.7 MD:61.7

5.1

1.0

ML:64.9 FL:66.2

1.3

3.3

Table 8.

SC

(higher is better) disparity on CelebSET. We reported the mean

SC

values for the two worst intersectional groups. In each mean column, left: worst, right: second-worst group.

WD_{g}

and

WD_{s}

denote Wasserstein distances between gender and skin-tone groups, respectively.

*

marks non-significant values (

p>0.05

). WD

>3

is highlighted. Values scaled by

100

	Accessories $\downarrow$		Hair $\downarrow$		Pose $\downarrow$		Multi-axis $\downarrow$
Model	Mean Values	$WD_{g}$	Mean Values	$WD_{g}$	Mean Values	$WD_{g}$	Mean Values	$WD_{g}$
Flux-Dev	M:82.5 F:57.1	25.3	M:85.0 F:51.6	33.5	M:65.3 F:33.8	31.5	M:84.4 F:50.0	34.4
Kontext-Pro	M:14.9 F:8.6	6.3	M:25.5 F:14.7	10.8	M:21.9 F:12.1	10.0	M:28.3 F:18.5	9.8
Nano Banana	M:15.0 F:6.8	8.3	M:18.0 F:12.6	5.3	F:10.5 M:9.8	1.0 ∗	M:21.9 F:14.2	7.7
Bagel-Edit	M:11.7 F:6.0	5.9	M:23.5 F:14.4	9.1	M:16.1 F:12.0	4.5	M:26.9 F:16.3	10.6
Qwen-Edit	M:16.2 F:11.7	4.5	M:30.4 F:20.4	10.0	M:26.4 F:14.7	11.7	M:33.3 F:22.4	11.1

Table 9. OER (lower is better) disparity on CelebA. We reported mean OER values for Male (M) and Female (F). In each mean column, left: worst, right: best group.

WD_{g}

denotes the Wasserstein distance between the male and female OER distributions.

*

marks non-significant values (

p>0.05

). WD

>3

is highlighted. Values are scaled by

100

	Accessories $\downarrow$			Hair $\downarrow$			Pose $\downarrow$			Multi-axis $\downarrow$
Model	Mean: 2 worst groups	$WD_{g}$	$WD_{s}$	Mean: 2 worst groups	$WD_{g}$	$WD_{s}$	Mean: 2 worst groups	$WD_{g}$	$WD_{s}$	Mean: 2 worst groups	$WD_{g}$	$WD_{s}$
Flux-Dev	MD:93.7 ML:69.9	23.1	19.9	MD:106.7 ML:76.9	35.1	24.5	MD:84.5 ML:56.0	32.4	24.4	MD:101.3 ML:81.1	36.8	22.3
Kontext-Pro	MD:25.5 ML:16.4	8.9	7.1	MD:38.5 FD:20.9	5.7	13.3	MD:19.0 ML:16.6	6.4	2.3 ∗	MD:41.4 ML:24.3	9.2	13.1
Nano Banana	FD:50.3 MD:32.9	8.5	22.0	MD:32.1 FD:15.3	6.6	13.8	MD:15.6 ML:9.2	3.9	3.7	MD:39.0 ML:21.5	9.8	11.4
Bagel-Edit	MD:19.9 FL:15.0	4.9	4.7	MD:38.7 FL:23.2	7.1	13.2	MD:17.3 ML:14.4	3.8	1.4 ∗	MD:42.8 ML:29.1	12.5	11.5
Qwen-Edit	MD:36.2 ML:22.2	9.0	9.2	MD:49.9 FL:28.5	6.5	14.8	MD:25.3 ML:21.5	10.3	2.4 ∗	MD:52.1 ML:31.6	11.8	15.3

Table 10. OER (lower is better) disparity on CelebSET. We reported the mean OER values for the two worst intersectional groups. In each mean column, left: worst, right: second-worst group.

WD_{g}

and

WD_{s}

denote Wasserstein distances between gender and skin-tone groups, respectively.

*

marks non-significant values (

p>0.05

). WD

>3

is highlighted. Values scaled by

100

Appendix figures & tables35 assets

Supplementary material from the paper’s appendix.

Appendix

Sub-axes	Categories	Attributes
	red hair	bright red hair; ruby red hair; intense scarlet red hair; crimson red hair; cherry red hair; rose red hair
	brown hair	light brown hair; light golden brown hair; honey brown hair; caramel brown hair; toffee brown hair; sandy brown hair; amber brown hair; ginger hazel brown hair; light hazel brown hair
color	black hair	black hair; jet black hair; raven black hair; ebony hair; charcoal black hair
	white hair	stark white hair; snow white hair; icy white hair; frost white hair; arctic white hair; nordic white hair
	blue hair	sky blue hair; aqua blue hair; electric blue hair; turquoise blue hair; teal blue hair; cobalt blue hair
	green hair	lime green hair; neon green hair; bright green hair; apple green hair; emerald green hair; spring green hair

Appendix

Table 11. Hair Attributes

Sub-axes	Categories	Attributes
glasses	round glasses	round glasses; round sunglasses; circular glasses; circular sunglasses; John Lennon glasses; oval glasses; oval sunglasses
	rectangle-shaped glasses	square glasses; square sunglasses; rectangular glasses; rectangular sunglasses; rectangle-shaped glasses; rectangle-shaped sunglasses
	baseball hat	baseball cap; baseball hat; snapback cap; snapback hat; fitted cap; fitted hat; dad hat
hats	beanie hat	beanie hat; knit hat; wool hat; slouchy beanie hat; winter beanie hat
	beret hat	beret hat; french beret hat
	panama hat	panama hat; straw panama hat; classic panama hat

Appendix

Table 12. Accessory Attributes

Sub-axes	Categories	Attributes
	looking straight ahead	looking straight ahead; frontal view; looking forward; head straight; front facing; looking ahead; head facing forward
head position	looking left	looking left; looking to the left; facing left; looking at the left; head facing left; head looking left; facing to the left; looking leftward; looking towards the left; looking left side
	looking right	looking right; looking to the right; facing right; looking at the right; head facing right; head looking right; facing to the right; looking rightward; looking towards the right; looking right side

Appendix

Table 13. Pose Attributes

Session	Step	Prompt
accessories_001	1	Add square glasses
	2	Add a wool hat
	3	Replace their wool hat with a knit hat.
	4	Change to rectangular sunglasses
accessories_002	1	Add a french beret hat.
	2	Add circular sunglasses.

Appendix

Table 14. Editing sessions for accessory-based edits

Session	Step	Prompt
hair_001	1	Add caramel brown hair
	2	Add pin-straight hair.
	3	Change the hair color to lavender hair.
	4	Add a bob cut.
hair_002	1	Add light hazel brown hair
	2	Change the hair color to neon peach hair.

Appendix

Table 15. Editing sessions for hair-based edits

Session	Step	Prompt
multi_axis_001	1	Make the person looking ahead.
	2	Add emerald green hair
	3	Add spiral curl.
	4	Have them looking straight ahead.
multi_axis_002	1	Add spiked hair.
	2	Add square glasses.

Appendix

Table 16. Editing sessions for multi-axis edits

Session	Step	Prompt
poses_001	1	Make them looking right.
	2	Change the head position to looking rightward.
	3	Change to head facing right
	4	Keep them facing right.
poses_002	1	Make them looking forward
	2	Make them looking to the right.

Appendix

Table 17. Editing sessions for pose-based edits

Panel of Figure 4 — Figure 4. Confusion matrices showing the alignment between human annotators and the Gemini-2.5-Flash MLLM.

Table 18. CelebA — number of edit sessions by gender across different edit types for each model

Table 19. CelebA — number of edit sessions by age across different edit types for each model

Table 20. CelebSET — number of edit sessions by gender and skin-tone across different edit types for each model.

Table 21. CelebSET — number of edit sessions by age across different edit types for each model

Figure 5. Non-significant p-val vs. WD scores: once WD > 3, almost all p-values are statistically significant.

	Accessories $\downarrow$		Hair $\downarrow$		Pose $\downarrow$		Multi-axis $\downarrow$
Model	Mean	$WD_{age}$	Mean	$WD_{age}$	Mean	$WD_{age}$	Mean	$WD_{age}$
Flux-Dev	O:90.2 Y:89.4	0.8	O:90.3 Y:88.7	1.6	O:90.3 Y:89.4	0.9	O:90.3 Y:89.1	1.2
Kontext-Pro	Y:26.0 O:23.5	2.4	O:25.6 Y:24.2	1.6	Y:38.3 O:34.9	3.4	Y:32.7 O:32.3	0.8
Nano Banana	Y:34.0 O:32.0	2.0	O:23.3 Y:23.0	0.5	Y:32.5 O:32.1	0.7 ∗	Y:32.0 O:32.0	0.5
Bagel-Edit	Y:27.9 O:25.7	2.3	O:19.3 Y:18.6	1.0	Y:54.1 O:49.6	4.5	Y:34.0 O:32.1	2.0
Qwen-Edit	Y:28.1 O:27.4	0.7	Y:28.0 O:28.0	1.7	Y:48.4 O:45.1	3.4	O:35.5 Y:35.0	1.4

Appendix

Table 22.

\operatorname{Drop}_{\mathrm{FSS}}

% (lower is better) disparity on CelebA. For each model and test set, we reported mean

\operatorname{Drop}_{\mathrm{FSS}}

for Young (Y) and Old (O) groups. In each mean column, the left value is the mean of the worst-performing group, and the right value is the mean of the best-performing group.

WD_{age}

denotes the Wasserstein distance between the young and old

\operatorname{Drop}_{\mathrm{FSS}}

distributions.

*

marks non-significant values (

p>0.05

). WD

>3

is highlighted as it indicates gender bias.

	Accessories $\downarrow$		Hair $\downarrow$		Pose $\downarrow$		Multi-axis $\downarrow$
Model	Mean	$WD_{age}$	Mean	$WD_{age}$	Mean	$WD_{age}$	Mean	$WD_{age}$
Flux-Dev	O:84.3 Y:83.4	1.5	O:83.8 Y:82.5	2.0	Y:83.2 O:82.8	1.9	O:83.3 Y:82.8	1.2
Kontext-Pro	O:27.9 Y:26.4	2.4	O:21.3 Y:21.3	2.7	Y:31.6 O:29.2	1.7	O:30.7 Y:29.5	3.4
Nano Banana	O:27.1 Y:25.3	3.8	O:17.7 Y:15.6	2.3	O:24.4 Y:21.6	3.2	O:25.1 Y:23.1	3.0
Bagel-Edit	O:28.9 Y:26.8	2.3	O:19.9 Y:17.3	2.8	Y:56.9 O:55.3	2.7	O:34.8 Y:33.8	2.7
Qwen-Edit	Y:48.7 O:44.0	5.1	Y:49.1 O:46.6	3.1	Y:41.8 O:37.7	3.4	Y:46.5 O:43.6	4.2

Appendix

Table 23.

\operatorname{Drop}_{\mathrm{FSS}}

% (lower is better) disparity on CelebSET. We reported the mean

\operatorname{Drop}_{\mathrm{FSS}}

for Young (Y) and Old (O) groups (left mean value: worst group; right mean value: best group).

WD_{age}

denotes the Wasserstein distance between the young and old

\operatorname{Drop}_{\mathrm{FSS}}

distributions.

*

marks non-significant values (

p>0.05

). WD

>3

is highlighted.

	Accessories $\uparrow$		Hair $\uparrow$		Pose $\uparrow$		Multi-axis $\uparrow$
Model	Mean	$WD_{age}$	Mean	$WD_{age}$	Mean	$WD_{age}$	Mean	$WD_{age}$
Flux-Dev	O:29.7 Y:35.1	5.4	O:26.9 Y:31.4	4.5	O:31.0 Y:31.5	0.9 ∗	O:27.4 Y:31.2	3.8
Kontext-Pro	Y:73.4 O:73.5	1.0 ∗	Y:75.9 O:76.9	1.1	O:46.6 Y:47.7	1.5 ∗	Y:63.9 O:64.3	0.6 ∗
Nano Banana	O:82.8 Y:85.4	2.6	O:86.9 Y:88.2	1.3	O:62.1 Y:62.8	0.9 ∗	O:73.8 Y:75.5	1.7
Bagel-Edit	O:87.1 Y:88.0	0.9	O:80.2 Y:80.3	0.3 ∗	O:56.8 Y:60.5	3.7	O:68.6 Y:69.0	0.8 ∗
Qwen-Edit	O:79.6 Y:80.0	0.7 ∗	Y:76.5 O:77.0	0.7 ∗	O:65.8 Y:68.1	2.3	O:63.8 Y:65.1	1.2

Appendix

Table 24.

SC

(higher is better) disparity on CelebA. We reported the mean

SC

for Young (Y) and Old (O) groups (left mean value: worst group; right mean value: best group).

WD_{age}

denotes the Wasserstein distance between the young and old

SC

distributions.

*

marks non-significant values (

p>0.05

). WD

>3

is highlighted. Values scaled by

100

	Accessories $\uparrow$		Hair $\uparrow$		Pose $\uparrow$		Multi-axis $\uparrow$
Model	Mean	$WD_{age}$	Mean	$WD_{age}$	Mean	$WD_{age}$	Mean	$WD_{age}$
Flux-Dev	O:26.4 Y:30.3	3.9	O:27.3 Y:28.8	1.6 ∗	O:28.5 Y:29.0	1.4 ∗	O:26.4 Y:27.4	1.3 ∗
Kontext-Pro	Y:72.5 O:73.5	1.6 ∗	O:74.5 Y:75.0	0.9 ∗	O:35.4 Y:40.5	5.1	O:59.9 Y:61.2	1.5 ∗
Nano Banana	Y:82.1 O:84.7	2.6	O:84.6 Y:86.8	2.2	O:58.7 Y:59.5	1.1 ∗	O:72.1 Y:73.4	1.5 ∗
Bagel-Edit	Y:86.2 O:88.5	2.2	Y:77.4 O:78.3	1.1 ∗	Y:59.9 O:61.1	1.5 ∗	O:67.5 Y:68.4	1.3 ∗
Qwen-Edit	O:81.9 Y:82.1	0.8 ∗	Y:77.2 O:78.7	1.5 ∗	O:60.8 Y:66.1	5.4	O:65.0 Y:66.4	1.7 ∗

Appendix

Table 25.

SC

(higher is better) disparity on CelebSET. We reported the mean

SC

for Young (Y) and Old (O) groups (left mean value: worst group; right mean value: best group).

WD_{age}

denotes the Wasserstein distance between the young and old

SC

distributions.

*

marks non-significant values (

p>0.05

). WD

>3

is highlighted. Values scaled by

100

	Accessories $\downarrow$		Hair $\downarrow$		Pose $\downarrow$		Multi-axis $\downarrow$
Model	Mean	$WD_{age}$	Mean	$WD_{age}$	Mean	$WD_{age}$	Mean	$WD_{age}$
Flux-Dev	O:78.6 Y:65.0	13.6	O:78.8 Y:62.3	16.5	O:59.6 Y:43.9	15.7	O:79.0 Y:60.8	18.3
Kontext-Pro	O:14.6 Y:10.3	4.3	O:24.8 Y:17.8	7.0	O:18.9 Y:15.7	3.4	O:28.0 Y:21.1	6.9
Nano Banana	O:14.1 Y:9.2	4.9	O:20.0 Y:13.4	6.7	O:13.5 Y:9.1	4.5	O:23.6 Y:15.6	8.0
Bagel-Edit	O:11.1 Y:7.6	3.7	O:24.4 Y:16.5	7.9	O:16.3 Y:13.0	3.6	O:25.9 Y:19.4	6.5
Qwen-Edit	O:16.6 Y:12.7	3.9	O:30.0 Y:23.2	6.9	O:20.4 Y:19.7	1.0 ∗	O:30.9 Y:26.1	4.8

Appendix

Table 26. OER (lower is better) disparity on CelebA. We reported the mean OER for the Young (Y) and Old (O) groups (left mean value: worst group; right mean value: best group).

WD_{age}

denotes the Wasserstein distance between the young and old OER distributions.

*

marks non-significant values (

p>0.05

). WD

>3

is highlighted. Values scaled by

100

	Accessories $\downarrow$		Hair $\downarrow$		Pose $\downarrow$		Multi-axis $\downarrow$
Model	Mean	$WD_{age}$	Mean	$WD_{age}$	Mean	$WD_{age}$	Mean	$WD_{age}$
Flux-Dev	O:81.3 Y:58.6	22.9	O:80.6 Y:58.5	22.1	O:62.6 Y:37.3	25.3	O:80.1 Y:56.5	23.8
Kontext-Pro	O:14.8 Y:13.0	3.1	O:26.3 Y:20.9	5.7	O:14.7 Y:13.7	2.4 ∗	O:29.9 Y:22.8	7.3
Nano Banana	Y:21.6 O:15.4	7.4	O:21.8 Y:12.5	9.2	O:9.2 Y:6.9	2.5 ∗	O:29.5 Y:24.1	10.1
Bagel-Edit	Y:12.7 O:11.8	5.1	O:26.2 Y:25.9	7.7	O:13.1 Y:12.5	2.3 ∗	O:34.5 Y:20.1	14.1
Qwen-Edit	O:25.0 Y:21.9	3.1 ∗	O:36.0 Y:29.9	6.4	O:23.4 Y:13.5	9.9	O:39.3 Y:29.6	9.8

Appendix

Table 27. OER (lower is better) disparity on CelebSET. We reported the mean OER for Young (age:

20-30

) and Old (age:

50-60

) groups (left mean value: worst group; right mean value: best group).

WD_{age}

denotes the Wasserstein distance between the young and old OER distributions.

*

marks non-significant values (

p>0.05

). WD

>3

is highlighted. Values scaled by

100

Panel of Figure 6 — Figure 6. $\operatorname{Drop}_{\mathrm{FSS}}$ $\%$ (lower is better) across gender subgroups in CelebA. Flux-Dev (F-Dev), Kontext-Pro (K-Pro), Nano Banana (NB), Bagel-Edit (BE), and Qwen-Edit (QE).

Panel of Figure 7 — Figure 7. $\operatorname{Drop}_{\mathrm{FSS}}$ $\%$ (lower is better) across intersectional demographic subgroups in CelebSET.

Panel of Figure 8 — Figure 8. $\operatorname{Drop}_{\mathrm{FSS}}$ $\%$ (lower is better) across age subgroups in CelebA.

Panel of Figure 9 — Figure 9. $\operatorname{Drop}_{\mathrm{FSS}}$ $\%$ (lower is better) across age subgroups in CelebSET.

Panel of Figure 10 — Figure 10. $SC$ (higher is better) across gender subgroups in CelebA.

Panel of Figure 11 — Figure 11. $SC$ (higher is better) across intersectional demographic subgroups in CelebSET.

Panel of Figure 12 — Figure 12. $SC$ (higher is better) across age subgroups in CelebA.

Panel of Figure 13 — Figure 13. $SC$ (higher is better) across age subgroups in CelebSET.

Panel of Figure 14 — Figure 14. OER (lower is better) across gender subgroups in CelebA.

Panel of Figure 15 — Figure 15. OER (lower is better) across intersectional demographic subgroups in CelebSET.

Panel of Figure 16 — Figure 16. OER (lower is better) across age subgroups in CelebA.

Panel of Figure 17 — Figure 17. OER (lower is better) across age subgroups in CelebSET.

Attribute	# Models	Model Names
spiked cut	5 (4M, 1F)	F-Dev, K-Pro, NB, BE, QE
straight hair	5 (5M)	F-Dev, K-Pro, NB, BE, QE
curly hair	4 (4M)	F-Dev, NB, BE, QE
fumanchu mustache	3 (3M)	F-Dev, K-Pro, NB
blue hair	3 (3M)	F-Dev, K-Pro, QE
mutton chops	3 (1M, 2F)	K-Pro, NB, BE

Appendix

Table 28. Attributes that frequently caused demographic biases in overediting. Model abbreviations: Nano-Banana (NB), SeedEdit (SE), Bagel-Edit (BE), Qwen-Edit (QE), Flux-Dev (F-Dev), and Kontext-Pro (K-Pro). Demographic abbreviations: Male (M), Female (F), Old (O), Young (Y), Light skin tone (L), and Dark skin tone (D).

Table 29. Baseline FSS (%) across demographic groups on CelebA and CelebSET. Subscripts denote

95\%

CI half-widths. ∗ superscript indicates a non-significant p-value (

p>0.05

). ML: Male-Light, FL: Female-Light, MD: Male-Dark, FD: Female-Dark.

Table 30. Stereotype analysis using

SC

. For the male-centric instruction subset, we reported the mean

SC

for demographic groups. In the Mean column, for CelebA: left is the worst group; right value is the best group. For CelebSET: left: worst group; right: second-worst group.

*

marks non-significant values (

p>0.05

). WD

>3

is highlighted. Values scaled by

100

Panel of Figure 18 — Figure 18. Facial similarity scores after each edit. The x-axis indicates the editing step, from $0$ (base image) to $4$ (final edited image). At each edit step, the Y-axis is the average facial similarity of a model for all images in that edit step. Note that $95\%$ confidence intervals are added, but are too narrow to be visible.

Text-Guided Image Editing

Text-guided image editing changes selected image content from natural-language instructions while aiming to preserve everything else.

See all 165 papers

RAGMesh with FaME-G2E: Long-Form Text-Driven 3D Face Generation and Editing
Aug 10, 2026Hao Li, Ju Dai, Feng Zhou +63D Mesh Generation 3D Shape Editing
State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing, China · Pengcheng Laboratory, Shenzhen, China · North China University of Technology, Beijing, China +3
FED-Bench: A Cross-Granular Benchmark for Disentangled Evaluation of Facial Expression Editing
Mar 31, 2026Fengjian Xue, Xuecheng Wu, Heli Sun +8Image Editing Image Editing Evaluation
Xi’an Jiaotong University · Northwestern Polytechnical University · National University of Singapore +2
Latent-Identity Tuning in Text-to-Image Personalization Models
Jul 13, 2026Daniel Garibi, Ronen Kamenetsky, Hadar Averbuch-Elor +2Subject-Driven Image Generation Identity-Preserving Image Generation
1Tel Aviv University · 2Cornell University

Aug 10, 2026J/K move · Enter open · S save

RAGMesh with FaME-G2E: Long-Form Text-Driven 3D Face Generation and Editing

Hao Li, Ju Dai, Feng Zhou, Mengting Shi+5

State Key Laboratory of Virtual Reality Technology and Systems, Beihang University, Beijing, China · Pengcheng Laboratory, Shenzhen, China · North China University of Technology, Beijing, China+3

3D Mesh Generation 3D Shape Editing Text-to-3D Generation

Open paper AskPDF arXiv

Text-driven 3D face generation and editing remains challenging due to the difficulty of translating long-form descriptions into fine-grained facial geometry. Existing methods primarily align global textual semantics with facial structures but often struggle to capture subtle local deformations, such as eyebrow tension, cheek contraction, and asymmetric mouth motions, resulting in limited geometric fidelity and editing precision. To facilitate fine-grained text-driven facial modeling, we first construct FaME-G2E, a large-scale multimodal dataset containing detailed text--mesh annotations and paired text--blendshape samples for unified 3D facial generation and editing. Based on this dataset, we propose RAGMesh, a retrieval-augmented framework that leverages text-correlated geometric priors to improve high-fidelity facial synthesis and editing. Specifically, the Multi-Scale Retrieval Fusion (MSRF) module retrieves semantically consistent global and regional facial priors and fuses them in the blendshape space, suppressing conflicting local deformations while preserving coherent deformation patterns. Furthermore, we introduce Adaptive RAG-guided Supervision (AdaRAGS), a region-aware constraint that explicitly aligns textual semantics with corresponding facial regions, enhancing regional controllability and editing accuracy. Extensive experiments on FaME-G2E demonstrate that RAGMesh achieves superior performance over state-of-the-art methods in local geometric accuracy, text-guided controllability, regional editing precision, and inference efficiency. Video demo is available at https://youtu.be/Yr0_XkpWcNk, and the source code and dataset will be released upon paper acceptance.

Abstract

Text-Guided Image Editing

Explore similar work

RAGMesh with FaME-G2E: Long-Form Text-Driven 3D Face Generation and Editing

FED-Bench: A Cross-Granular Benchmark for Disentangled Evaluation of Facial Expression Editing

Latent-Identity Tuning in Text-to-Image Personalization Models