Traversing the Satisfaction-Diversity Frontier in Text-to-Image Diffusion

Kevin Zhai; Siva Rajesh Kasa; Soumya Roy; Sumit Negi; Mubarak Shah

cs.AIOct 1, 2026

Traversing the Satisfaction-Diversity Frontier in Text-to-Image Diffusion

Authors: Kevin Zhai, Siva Rajesh Kasa, Soumya Roy, Sumit Negi, Mubarak Shah

Organizations: University of Central Florida · Amazon

Abstract

Text-to-image generation enables users to explore several images generated from the same prompt. For these generated images to be useful, each one must reflect the user's preferences, measured by a learned reward, and differ visually from the others to maintain diversity. Existing methods are limited: they either address reward and diversity separately or combine them in one aggregate score, enabling high diversity to offset low rewards. In this paper, we address these limitations by formulating generation as satisficing: every image (candidate) must satisfy a reward floor and the batch of images must satisfy a diversity cutoff. The reward floor controls the balance between worst-candidate reward and batch diversity; we show that varying this floor defines a Pareto frontier. To traverse this frontier, we introduce SatisDive, a training-free inference-time method. SatisDive uses a batch-relative reward cutoff to distinguish lower- from higher-reward candidates, emphasizing reward improvement for candidates below the cutoff and diversity among candidates above it. On Pick-a-Pic, at matched DreamSim, SatisDive improves worst-candidate reward over FK steering by up to 0.43 with FLUX.1-dev as the base model and HPSv3 as the reward, and by up to 0.70 with SANA-1.6B as the base model and ImageReward as the reward. More broadly, across their overlapping DreamSim ranges, SatisDive's satisfaction-diversity curve Pareto-dominates FK steering's curve in each setting.

Figures & tables

Figure 1: SatisDive returns four visually distinct images above the reward floor, while FK steering returns four nearly identical images. We use the SANA-1.6B ( Xie et al., 2025 ) flow matching model with ImageReward ( Xu et al., 2023 ) reward model. On a Pick-a-Pic ( Kirstain et al., 2023 ) prompt depicting two colliding viruses, the minimum reward is $0.66$ for FK steering and $1.25$ for SatisDive, with reward floor $\rho{=}1.19$ .

Figure 2: In a toy setting, each baseline tends towards high diversity or high reward, not both. Each method generates $K{=}4$ candidates from a frozen flow model trained on a mixture of Gaussians (details in Appendix C.1 ). Reward is bounded in $(0,1]$ , and the midpoint $\rho=0.5$ is a reward floor that every candidate must satisfy. (a) To visualize clearly the coverage of high-reward regions, candidates are 2D points on a Cartesian $x$ – $y$ grid rather than images. We show one selected illustrative batch of $K{=}4$ candidates per method. DAS’s candidates are all above $\rho$ but are not diverse; they cover only a single high-reward region (yellow circles). Conversely, NegToMe’s candidates are diverse yet are all below $\rho$ (red circles). In these batches, only ours has all candidates above $\rho$ in distinct high-reward regions. (b) We plot worst-candidate reward against diversity (mean pairwise distance) averaged over 200 runs for each method. Methods in the shaded green region have average worst-candidate rewards greater than $\rho$ . DAS achieves the highest reward but lowest diversity; NegToMe achieves the highest diversity but lowest reward.

Panel of Figure 3 — Figure 3: SatisDive achieves better worst-candidate reward at matched diversity versus FK steering. We plot mean worst-candidate reward $\min_{k}r_{k}$ against mean pairwise DreamSim, on Pick-a-Pic. We sweep SatisDive and FK steering on two settings. (a) FLUX.1-dev is the base model and HPSv3 is the reward, with $\Delta$ from $0.0001$ to $3$ and $\lambda$ from $0.5$ to $2$ . (b) SANA-1.6B is the base model and ImageReward is the reward, with $\Delta$ from $0.1$ to $1.5$ and $\lambda$ from $0.5$ to $40$ . The shaded band represents the worst-candidate reward gap between the two curves at the same diversity, marked at $\Delta{=}1.5$ in (a) where it is $\approx\!0.43$ on FLUX HPSv3, and at $\Delta{=}0.5$ in (b) where it is $\approx\!0.70$ on SANA ImageReward. We note that the top right region, representing high reward and high diversity, is ideal. DAS and VASR collapse to diversity below $0.06$ in both settings, and NegToMe is the most diverse but its worst candidate on average falls below the base model’s. SatisDive traces a better empirical frontier than FK steering between these extremes.

	Satisfaction–diversity			Additional metrics		Held-out
Method	min $\uparrow$	DreamSim $\uparrow$	$\mathrm{SAT}_{\rho}$ $\uparrow$	mean $\uparrow$	LPIPS $\uparrow$	ImageReward	PickScore
Base	$8.60$	$0.663$ $\,\pm 0.013$	$0.500$ $\,\pm 0.038$	$9.65$	$0.643$ $\,\pm 0.008$	$1.02$	$22.28$
NegToMe	$6.30$	$0.885$ $\,\pm 0.009$	$0.276$ $\,\pm 0.027$	$8.17$	$0.788$ $\,\pm 0.004$	$0.79$	$21.49$
FK steering	$9.54$	$0.306$ $\,\pm 0.021$	$0.570$ $\,\pm 0.041$	$10.07$	$0.291$ $\,\pm 0.021$	$1.09$	$22.37$
FK + NegToMe	$7.69$	$0.360$ $\,\pm 0.017$	$0.333$ $\,\pm 0.034$	$8.70$	$0.310$ $\,\pm 0.016$	$0.99$	$21.68$
VASR	$9.99$	$0.053$ $\,\pm 0.004$	$0.598$ $\,\pm 0.042$	$10.17$	$0.033$ $\,\pm 0.005$	$1.10$	$22.40$

Table 1: On Pick-a-Pic, SatisDive is competitive on satisfaction rate without collapsing diversity. We use FLUX.1-dev with HPSv3 as the reward. DAS has the highest satisfaction rate at

0.612

but a mean pairwise DreamSim of only

0.059

, while NegToMe is the most diverse at

0.885

and has the lowest satisfaction rate at

0.276

. SatisDive attains a satisfaction rate of

0.552

at a DreamSim of

0.555

, and matches base FLUX on both held-out rewards. We use

\rho{=}9.66

;

\pm

denotes 95% CIs.

	Satisfaction–diversity			Additional metrics		Held-out
Method	min $\uparrow$	DreamSim $\uparrow$	$\mathrm{SAT}_{\rho}$ $\uparrow$	mean $\uparrow$	LPIPS $\uparrow$	ImageReward	PickScore
Base	$12.34$	$0.655$ $\,\pm 0.004$	$0.500$ $\,\pm 0.015$	$13.36$	$0.646$ $\,\pm 0.003$	$1.11$	$22.73$
NegToMe	$9.65$	$0.865$ $\,\pm 0.003$	$0.309$ $\,\pm 0.011$	$12.02$	$0.781$ $\,\pm 0.002$	$0.91$	$21.95$
FK steering	$13.44$	$0.197$ $\,\pm 0.006$	$0.565$ $\,\pm 0.016$	$13.77$	$0.182$ $\,\pm 0.007$	$1.17$	$22.79$
FK + NegToMe	$11.53$	$0.312$ $\,\pm 0.005$	$0.320$ $\,\pm 0.013$	$12.50$	$0.277$ $\,\pm 0.005$	$1.06$	$22.04$
VASR	$13.62$	$0.053$ $\,\pm 0.002$	$0.569$ $\,\pm 0.017$	$13.78$	$0.033$ $\,\pm 0.002$	$1.17$	$22.79$

Table 2: On HPDv2, SatisDive is competitive on satisfaction rate without collapsing diversity. We use FLUX.1-dev with HPSv3 as the reward. SatisDive attains a satisfaction rate of

0.553

at a mean pairwise DreamSim of

0.567

, against DAS’s

0.580

0.082

, and matches base FLUX on both held-out rewards.

\mathrm{SAT}_{\rho}

uses floor

\rho{=}13.93

\pm

denotes 95% confidence intervals.

	Satisfaction–diversity			Additional metrics		Held-out
Method	min $\uparrow$	DreamSim $\uparrow$	$\mathrm{SAT}_{\rho}$ $\uparrow$	mean $\uparrow$	LPIPS $\uparrow$	HPSv3	PickScore
Base	$0.41$	$0.706$ $\,\pm 0.012$	$0.500$ $\,\pm 0.037$	$0.91$	$0.644$ $\,\pm 0.010$	$7.53$	$21.93$
NegToMe	$-0.29$	$0.879$ $\,\pm 0.011$	$0.380$ $\,\pm 0.033$	$0.55$	$0.743$ $\,\pm 0.009$	$5.98$	$21.29$
FK steering	$1.06$	$0.254$ $\,\pm 0.022$	$0.625$ $\,\pm 0.041$	$1.18$	$0.204$ $\,\pm 0.021$	$7.98$	$22.06$
FK + NegToMe	$0.84$	$0.386$ $\,\pm 0.025$	$0.585$ $\,\pm 0.041$	$1.04$	$0.278$ $\,\pm 0.022$	$7.07$	$21.63$
VASR	$1.23$	$0.046$ $\,\pm 0.003$	$0.651$ $\,\pm 0.041$	$1.25$	$0.023$ $\,\pm 0.003$	$8.01$	$22.09$

Table 3: SatisDive exceeds every reward-steering method on worst-candidate reward and diversity on SANA-1.6B. We use Pick-a-Pic with ImageReward; the reward floor is

\rho\!=\!1.19

. Our method is shown at three

\Delta

settings, tracing its satisfaction–diversity frontier as reflected in Figure 3 b. At

\Delta=0.3

, SatisDive attains a worst-candidate reward of

1.29

and DreamSim of

0.591

, exceeding every reward-steering method on both metrics.

\pm

denotes 95% confidence intervals.

Figure 4: On two Pick-a-Pic prompts, only SatisDive returns four visually distinct images that are all above the reward floor. We use SANA-1.6B with ImageReward and matched starting seeds; the reward floor is $\rho{=}1.19$ . The magenta border marks the worst-reward candidate in each row, whose reward is reported on the right. Base and NegToMe produce visually varied batches, but their worst-reward candidates do not satisfy the reward floor. NegToMe’s worst-reward car image (on the left) is nearly black, and its worst-reward fairy image (on the right) does not clearly show an owl. FK steering’s images exceed the reward floor, but its four images repeat nearly the same composition.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Description
$c$	Text prompt
$K$	Number of candidates in a batch
$k$	Candidate index
$T$	Number of sampling steps
$s$	Discrete denoising-step index, $s=1,\ldots,T$
$t$	Model time, decreasing from $1$ (noise) to $0$ (image)

Appendix

Table 4: Summary of Notation.

	Satisfaction–diversity			Additional metric	Held-out
Method	min $\uparrow$	DreamSim $\uparrow$	$\mathrm{SAT}_{\rho}$ $\uparrow$	LPIPS $\uparrow$	HPSv3	PickScore
Base	$-1.19$	$0.89$	$0.50$	$0.76$	$5.74$	$20.49$
NegToMe	$-1.82$	$1.01$	$0.30$	$0.83$	$2.00$	$19.42$
FK steering	$0.69$	$0.45$	$0.88$	$0.37$	$7.25$	$20.91$
FK + NegToMe	$-1.03$	$0.64$	$0.46$	$0.53$	$0.43$	$19.26$
VASR	$0.74$	$0.47$	$0.88$	$0.38$	$7.55$	$20.96$

Appendix

Table 5: At

K{=}16

, SatisDive maintains a competitive satisfaction rate and higher diversity than the reward-steering baselines. We use SD1.5 with ImageReward,

\Delta{=}0.75

, and

\rho{=}0.16

, the median reward of the base model’s

K{=}16

outputs.

	Satisfaction–diversity		Additional metric
Method	DreamSim $\uparrow$	$\mathrm{SAT}_{\rho}$ $\uparrow$	LPIPS $\uparrow$
Base	$0.70$	$0.50$	$0.64$
NegToMe	$0.90$	$0.34$	$0.78$
FK steering	$0.18$	$0.57$	$0.16$
FK + NegToMe	$0.31$	$0.37$	$0.27$
VASR	$0.05$	$0.56$	$0.03$

Appendix

Table 6: SatisDive has the highest observed satisfaction rate on DrawBench and higher observed diversity than every reward-steering method. We use FLUX.1-dev with HPSv3 on the

200

DrawBench prompts and set

\rho{=}11.60

Figure 5: At

\Delta\in\{0.1,0.3,0.5,0.75\}

, SatisDive has a higher observed satisfaction rate and DreamSim than every reward-steering baseline under the stricter floor. We use SANA-1.6B with ImageReward and set

\rho{=}1.69

, the

75

th percentile of the base model’s candidate rewards. Error bars are

95\%

bootstrap confidence intervals over prompts. For legibility, we show five of the seven

\Delta

settings.

Method	CMMD $\downarrow$
Base	$0.87$
NegToMe	$0.83$
FK steering	$0.88$
DAS	$0.87$
Ours	$\mathbf{0.81}$

Appendix

Table 7: SatisDive has the lowest observed CMMD among the evaluated methods. We report CMMD (

\times 10^{3}

; lower is better) using

2000

images generated from

500

MS-COCO captions per method and all

5000

MS-COCO val2017 images as the reference distribution. We use FLUX.1-dev with HPSv3,

\Delta{=}1.5

for SatisDive,

\lambda{=}1.5

for FK steering,

\alpha{=}0.1

and

\gamma{=}0.024

for DAS, and the published NegToMe settings.

	Satisfaction–diversity			Additional metric	Held-out
Variant	min $\uparrow$	DreamSim $\uparrow$	$\mathrm{SAT}_{\rho}$ $\uparrow$	mean $\uparrow$	ImageReward	PickScore
Full method	$9.39$	$0.56$	$0.55$	$10.03$	$1.04$	$22.32$
w/o replacement	$8.58$	$0.68$	$0.50$	$9.65$	$1.01$	$22.27$
w/o diversity penalty	$9.43$	$0.54$	$0.56$	$10.07$	$1.07$	$22.35$

Appendix

Table 8: Removing late-stage latent replacement lowers worst-candidate reward, while removing the diversity penalty lowers DreamSim. We remove one component at a time from SatisDive using FLUX.1-dev with HPSv3,

\Delta{=}1.5

, and

\rho{=}9.66

. The

95\%

bootstrap confidence intervals computed from per-prompt differences exclude zero for the changes in worst-candidate and mean HPSv3 after removing late-stage latent replacement and for the changes in DreamSim and mean HPSv3 after removing the diversity penalty.

	Satisfaction–diversity			Held-out
Variant	min $\uparrow$	DreamSim	$\mathrm{SAT}_{\rho}$ $\uparrow$	HPSv3	PickScore
Full method	$0.54$ $\,\pm 0.08$	$0.72$	$0.77$	$6.40$	$20.69$
w/o replacement	$-0.07$	$0.88$	$0.64$	$5.94$	$20.57$

Appendix

Table 9: late-stage latent replacement also improves worst-candidate reward and satisfaction on Stable Diffusion 1.5. We compare the full method with a variant without latent replacement using Stable Diffusion 1.5 with ImageReward on Pick-a-Pic,

\Delta{=}0.75

, and

\rho{=}0.18

. The

\pm

value is the half-width of a

95\%

prompt-bootstrap confidence interval, where available.

	Satisfaction–diversity			Additional metric	Held-out
Method	min $\uparrow$	DreamSim $\uparrow$	$\mathrm{SAT}_{\rho}$ $\uparrow$	LPIPS $\uparrow$	HPSv3	PickScore
Base	$0.68$	$0.66$ $\,\pm 0.01$	$0.50$ $\,\pm 0.04$	$0.64$	$9.65$	$22.28$
NegToMe	$0.20$	$0.88$ $\,\pm 0.01$	$0.39$ $\,\pm 0.03$	$0.79$	$8.17$	$21.49$
FK steering	$1.16$	$0.23$ $\,\pm 0.02$	$0.61$ $\,\pm 0.04$	$0.20$	$9.74$	$22.38$
FK + NegToMe	$0.97$	$0.45$ $\,\pm 0.02$	$0.56$ $\,\pm 0.04$	$0.38$	$7.45$	$21.33$
VASR	$1.24$	$0.05$ $\,\pm 0.00$	$0.62$ $\,\pm 0.04$	$0.03$	$9.79$	$22.40$

Appendix

Table 10: With ImageReward, SatisDive matches FK steering’s satisfaction while maintaining higher diversity. We repeat the FLUX.1-dev comparison using ImageReward,

\Delta{=}0.5

, and

\rho{=}1.23

. The

\pm

values are half-widths of

95\%

percentile bootstrap confidence intervals over prompts.

	Satisfaction–diversity			Additional metric	Held-out
Method	min $\uparrow$	DreamSim $\uparrow$	$\mathrm{SAT}_{\rho}$ $\uparrow$	LPIPS $\uparrow$	HPSv3	PickScore
Base	$-0.62$	$0.89$ $\,\pm 0.01$	$0.50$ $\,\pm 0.03$	$0.77$	$5.77$	$20.48$
NegToMe	$-1.24$	$1.03$ $\,\pm 0.01$	$0.32$ $\,\pm 0.03$	$0.84$	$2.79$	$19.63$
FK steering	$0.35$	$0.49$ $\,\pm 0.03$	$0.72$ $\,\pm 0.04$	$0.40$	$6.71$	$20.78$
FK + NegToMe	$-0.63$	$0.71$ $\,\pm 0.03$	$0.45$ $\,\pm 0.04$	$0.56$	$1.63$	$19.53$
VASR	$0.62$	$0.36$ $\,\pm 0.01$	$0.77$ $\,\pm 0.04$	$0.26$	$7.00$	$20.81$

Appendix

Table 11: SatisDive achieves higher satisfaction and diversity than FK steering on Stable Diffusion 1.5. We use ImageReward with

\rho{=}0.18

and report SatisDive at three values of

\Delta

. The

\pm

values are half-widths of

95\%

percentile bootstrap confidence intervals over prompts.

	Satisfaction–diversity			Additional metric
Method	min $\uparrow$	DreamSim $\uparrow$	$\mathrm{SAT}_{\rho}$ $\uparrow$	mean $\uparrow$
CREPE (scale $10$ )	$-0.33$ $\,\pm 0.09$	$0.761$ $\,\pm 0.011$	$0.528$ $\,\pm 0.036$	$0.20$ $\,\pm 0.08$
CREPE (scale $100$ )	$0.07$ $\,\pm 0.08$	$0.753$ $\,\pm 0.011$	$0.660$ $\,\pm 0.034$	$0.52$ $\,\pm 0.07$
Ours ( $\Delta{=}0.5$ )	$0.70$ $\,\pm 0.08$	$0.650$ $\,\pm 0.021$	$0.802$ $\,\pm 0.033$	$0.88$ $\,\pm 0.08$
Ours ( $\Delta{=}1.0$ )	$0.44$ $\,\pm 0.08$	$0.762$ $\,\pm 0.016$	$0.740$ $\,\pm 0.034$	$0.71$ $\,\pm 0.08$
Ours ( $\Delta{=}1.25$ )	$0.32$ $\,\pm 0.09$	$0.796$ $\,\pm 0.015$	$0.708$ $\,\pm 0.035$	$0.63$ $\,\pm 0.08$

Appendix

Table 12: SatisDive improves worst-candidate reward at comparable diversity to both evaluated CREPE settings. Both methods return

K{=}4

images per prompt using SD1.5 with ImageReward. We use

\rho{=}0.18

and report means on Pick-a-Pic;

\pm

values are half-widths of

95\%

percentile confidence intervals from

200{,}000

prompt-bootstrap resamples, keeping each prompt’s four candidates together and

\rho

fixed.

$\lambda_{\mathrm{SGI}}$	min $\uparrow$	DreamSim $\uparrow$	$\mathrm{SAT}_{\rho}$ $\uparrow$	mean $\uparrow$
	Satisfaction–diversity			Additional metric
$0$	$9.32$ $\,\pm 0.20$	$0.647$ $\,\pm 0.012$	$0.592$ $\,\pm 0.037$	$10.20$
$1.5$	$9.17$ $\,\pm 0.20$	$0.693$ $\,\pm 0.012$	$0.577$ $\,\pm 0.037$	$10.15$
$6$	$8.89$ $\,\pm 0.20$	$0.744$ $\,\pm 0.013$	$0.563$ $\,\pm 0.036$	$10.04$
$24$	$8.41$ $\,\pm 0.22$	$0.803$ $\,\pm 0.013$	$0.529$ $\,\pm 0.035$	$9.82$
$96$	$8.12$ $\,\pm 0.23$	$0.823$ $\,\pm 0.013$	$0.507$ $\,\pm 0.035$	$9.65$

Appendix

Table 13: Increasing SGI’s diversity weight increases DreamSim but decreases satisfaction rate. We evaluate SGI on Pick-a-Pic using FLUX.1-dev with HPSv3. SGI prunes

M{=}128

candidates to

K{=}4

, and

\lambda_{\mathrm{SGI}}

weights its DINO diversity score. We use

\rho{=}9.66

; the

\pm

values are half-widths of

95\%

bootstrap confidence intervals over prompts.

Figure 6: Inference-time operation counts for the evaluated FLUX.1-dev methods. One image-equivalent pass processes one candidate once through the indicated model; processing a batch of

b

candidates therefore counts as

b

passes. We count operations per prompt for

K{=}4

returned candidates and

28

denoising steps. (a) SatisDive uses at most

200

diffusion-model passes, compared with

112

for Base, NegToMe, FK steering, FK + NegToMe, and VASR,

224

for DAS, and

340

for SGI. The SatisDive count is an upper bound because

24

of its

48

backward passes stop at the intermediate feature capture. DAS computes a reward gradient at every denoising step; its SSP resampling condition is checked at steps

4,8,\ldots,24

. (b) We count auxiliary-model forward evaluations. SGI uses

252

HPSv3 evaluations and

252

DINO evaluations while pruning

M{=}128

candidates to

K{=}4

. The counts exclude VAE decoding and SGI’s candidate-selection optimization.

Method	Time (s/prompt)	Total GPU memory (GiB)	GPUs used
Base	22.5	42.2	1
NegToMe	25.9	42.2	1
FK steering	37.7	58.5	2
FK + NegToMe	40.5	58.5	2
VASR	37.6	58.5	2
DAS	117.2	78.2	2

Appendix

Table 14: Measured runtime and GPU memory for FLUX.1-dev. Each prompt returns

K{=}4

images at

1024\times 1024

resolution using

28

denoising steps and HPSv3 for reward-guided methods. Time is the mean over three prompts, excluding model loading and image-file writes. Memory is the largest sampled resident GPU memory, summed across GPUs over generation and reward-model processes, rather than an exact peak. All runs use H100 80 GB GPUs; two-GPU methods place the generator and reward model on separate devices. Figure 3 reports the corresponding worst-candidate reward improvements at matched diversity.

	Satisfaction–diversity			Additional metric	Held-out
Run	min $\uparrow$	DreamSim $\uparrow$	$\mathrm{SAT}_{\rho}$ $\uparrow$	LPIPS $\uparrow$	HPSv3	PickScore
Seed $0$	$0.54$ $\,\pm 0.08$	$0.72$ $\,\pm 0.02$	$0.77$ $\,\pm 0.03$	$0.64$	$6.40$	$20.69$
Seed $1$	$0.59$ $\,\pm 0.08$	$0.72$ $\,\pm 0.02$	$0.76$ $\,\pm 0.03$	$0.64$	$6.51$	$20.71$
Seed $2$	$0.59$ $\,\pm 0.08$	$0.71$ $\,\pm 0.02$	$0.76$ $\,\pm 0.03$	$0.63$	$6.52$	$20.69$
Seed $3$	$0.58$ $\,\pm 0.08$	$0.71$ $\,\pm 0.02$	$0.78$ $\,\pm 0.03$	$0.62$	$6.63$	$20.72$
Seed $4$	$0.54$ $\,\pm 0.08$	$0.73$ $\,\pm 0.02$	$0.75$ $\,\pm 0.03$	$0.65$	$6.42$	$20.69$

Appendix

Table 15: SatisDive’s satisfaction and diversity remain stable across five initial-noise seeds. We use Stable Diffusion 1.5 with ImageReward on the same Pick-a-Pic prompts, set

\Delta{=}0.75

, and compute

\beta_{r}

separately for each prompt and seed. The

\pm

values are half-widths of

95\%

bootstrap confidence intervals over prompts.

	Satisfaction–diversity			Additional metric	Held-out
Method	min $\uparrow$	DreamSim $\uparrow$	$\mathrm{SAT}_{\rho}$ $\uparrow$	mean $\uparrow$	HPSv3
DAS	$1.10$ $\,\pm 0.07$	$0.016$ $\,\pm 0.001$	$0.599$ $\,\pm 0.043$	$1.11$	$7.72$
DAS (reward capped)	$1.06$ $\,\pm 0.07$	$0.066$ $\,\pm 0.008$	$0.577$ $\,\pm 0.042$	$1.09$	$7.65$
Ours ( $\Delta{=}0.75$ )	$1.13$ $\,\pm 0.07$	$0.658$ $\,\pm 0.015$	$0.674$ $\,\pm 0.039$	$1.28$	$7.63$

Appendix

Table 16: Reward-capped DAS remains less diverse than SatisDive. We use SANA-1.6B with ImageReward on Pick-a-Pic and set

\rho{=}1.19

. The reward-capped variant replaces

r_{k}

with

\min(r_{k},\rho)

in the guidance gradient and resampling weights. The

\pm

values are half-widths of

95\%

bootstrap confidence intervals over prompts.

Figure 7: On the Japanese-highway car prompt, SatisDive returns four visually distinct candidates above the reward floor. SatisDive attains a minimum ImageReward of $1.64$ and a DreamSim of $0.73$ . FK steering, VASR, DAS, and FK + NegToMe also satisfy the reward floor, but their DreamSim values are $0.11$ , $0.10$ , $0.02$ , and $0.22$ . Base and NegToMe remain diverse, but their minimum rewards of $1.13$ and $-0.03$ fall below the reward floor.

Figure 8: On the fairy-and-owl prompt, SatisDive has the highest minimum ImageReward, $1.73$ , and a DreamSim of $0.68$ . FK steering, VASR, DAS, and FK + NegToMe have DreamSim values between $0.02$ and $0.23$ . Base and NegToMe remain diverse, but their minimum rewards of $0.69$ and $-1.44$ are below the reward floor.

Figure 9: On the dog-driving-a-bus prompt, SatisDive attains a minimum ImageReward of $1.46$ and a DreamSim of $0.84$ . VASR and DAS satisfy the reward floor but return visually similar candidates. Base and NegToMe have comparable diversity to SatisDive, but each includes candidates below the reward floor.

Figure 10: For Stable Diffusion 1.5, SatisDive returns four visually distinct images with a minimum ImageReward of $1.72$ . The prompt describes a man on a rocky precipice above clouds. Base and NegToMe have minimum rewards of $0.49$ and $0.03$ , while the candidates from FK steering, VASR, and DAS are visually similar within each row. All methods use matched starting seeds.

T2I Generation

Text-to-image generation creates images from natural-language prompts using generative models, with research focused on fidelity, control, efficiency, and safety.

See all 303 papers

Post-Training Frontier Text-to-Image Models by Composing Preference and Rubric Rewards
Oct 2, 2026Yuanhao Ban, I-Hung Hsu, Anastasios Angelopoulos +3T2I Generation Reward Design for RL
Arena Intelligence Inc · UCLA
TPSO: Training-Free Diverse Image Generation via Semantic Prompt Embedding Optimization
Nov 25, 2025Debin Meng, Chen Jin, Zheng Gao +3T2I Generation T2I Diffusion Models
Queen Mary University of London · Centre for AI, AstraZeneca, UK · University of Bedfordshire +1
Breaking the Lock-in: Diversifying Text-to-Image Generation via Representation Modulation
Jun 5, 2026Dahee Kwon, Haeun Lee, Jaesik ChoiDiffusion Transformer T2I Generation
Korea Advanced Institute of Science and Technology (KAIST) · INEEJI

Oct 2, 2026J/K move · Enter open · S save

Post-Training Frontier Text-to-Image Models by Composing Preference and Rubric Rewards

Yuanhao Ban, I-Hung Hsu, Anastasios Angelopoulos, Wei-Lin Chiang+2

Arena Intelligence Inc · UCLA

T2I Generation Reward Design for RL RL for Image Generation Reward Modeling T2I Generation Evaluation

Open paper AskPDF arXiv

Recent text-to-image generation models have achieved remarkable visual quality, but improving them through post-training remains challenging because no single reward signal captures the full range of human preference. In this work, we develop a simple and effective post-training recipe for open-domain text-to-image generation based on the composition of complementary reward signals. Our reward system consists of two main components: a preference reward, trained on large-scale human preference data using a Bradley-Terry objective to capture overall human aesthetic and perceptual preferences, and rubric-based rewards, which explicitly evaluate prompt faithfulness and other desirable properties while providing safeguards against reward hacking. A key challenge is how to combine these heterogeneous reward signals. We show that a naive weighted average leads to suboptimal optimization behavior, and propose a simple reward composition strategy that more effectively balances preference optimization with rubric satisfaction. In the Arena text-to-image leaderboard (https://arena.ai/), our RL-trained Flux2dev achieves an Elo rating 69 points above the base model, and our post-trained Ideogram-4 surpasses every open-source model on the leaderboard, reaching an Elo of 1223.5. (Claims of state-of-the-art performance are based on the Arena leaderboard snapshot as of September 4, 2026.) Our results suggest that effective rewards for frontier generative-model training require broad coverage of user intent and robustness to exploitation under optimization. To support reproducible research, we release Arena-T2I-Training, a 1K subset of training data that recovers some gains of full-scale training, providing a resource that we hope will facilitate future work on post-training for text-to-image models.

Abstract

T2I Generation

Explore similar work

Post-Training Frontier Text-to-Image Models by Composing Preference and Rubric Rewards

TPSO: Training-Free Diverse Image Generation via Semantic Prompt Embedding Optimization

Breaking the Lock-in: Diversifying Text-to-Image Generation via Representation Modulation