Semantic-First: Temporal Scheduling in Flow Matching for Reliable Emotional Image Generation

Zhenxuan Wang

cs.CVOct 24, 2026

Semantic-First: Temporal Scheduling in Flow Matching for Reliable Emotional Image Generation

Authors: Zhenxuan Wang

Organizations: Lanzhou University

Abstract

Affect-controlled image generation should change how a scene feels while preserving what the prompt specifies. We study the temporal allocation of stochastic exploration and policy optimization in flow-matching models as a means of controlling this balance. During RL training, our semantic-first schedule uses deterministic ordinary differential equation (ODE) sampling during the high-noise phase and confines stochastic differential equation (SDE) exploration and group relative policy optimization (GRPO) to a late window. Controlled experiments separate exploration location, optimization location, and terminal-step sampling, showing distinct semantic benefits from late exploration and late updates. On FLUX.1-dev, semantic-first improves prompt faithfulness, independently evaluated affect accuracy, and human-rated semantic preservation and overall preference. The expanded semantic audit records 2.7% subject drift, while GRPO training time decreases by 41.1% relative to original full-path Flow-GRPO. Window-width experiments characterize the affect-semantic and computational trade-offs, and replication on SD3.5-Medium reproduces the scheduling advantage. These results establish denoising time as a practical control dimension for reliable emotional image generation.

Figures & tables

Figure 1: With the same prompt, V–A target and initial noise, early exploration introduces a foreground traveler; late exploration retains the mountain landscape and night sky. Percentages are aggregate drift rates from the 150-image-per-method audit (Table 5 ). Vector-layout code: OpenAI Codex (desktop 26.1021.3705.0), reviewed by the author.

Figure 2: Semantic-first training. (a) The residual EIT maps neutral-prompt T5 features and a continuous V–A target to fixed FLUX conditions. (b) ODE steps 0–19 precede SDE exploration and GRPO loss at 20–23, followed by ODE readout at 24. The trained policy serves all inference steps. (c) Fixed/progressive schedules use four stochastic transitions; A uses an SDE terminal step and B uses ODE. Controls: Table 2 . Vector-layout code: OpenAI Codex (desktop 26.1021.3705.0), reviewed by the author.

Figure 3: Pilot trends as the SDE window moves from high-noise to low-noise steps. (a) V/A errors in the original

[-3,3]

coordinate units; (b) CLIPScore, equal to cosine similarity multiplied by 100; (c) TIFA and DINO on their 0–100 presentation scales. Points show means and error bars show sample SD over training seeds 42, 123, and 3407, with 200 updates per run.

Method	V-Error $\downarrow$	A-Error $\downarrow$	CLIP $\uparrow$	DINO $\uparrow$	IQA $\uparrow$	TIFA $\uparrow$
EmotiCrafter	$1.510$	$1.828$	$27.74$	$77.3$	$0.917$	$74.82$
FLUX+VA-EIT	$1.402\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.048$	$1.710\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.036$	$29.27\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.27$	$\mathbf{81.6}\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.9$	$\mathbf{0.924}\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.024$	$80.49\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.63$
A: Full-path	$\mathbf{0.947}\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.046$	$\mathbf{1.191}\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.037$	$27.34\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.19$	$64.7\mskip 2.0mu\mathord{\pm}\mskip 2.0mu6.2$	$0.907\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.021$	$79.75\mskip 2.0mu\mathord{\pm}\mskip 2.0mu1.08$
B: Full-path	$0.963\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.041$	$1.208\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.034$	$27.58\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.21$	$65.5\mskip 2.0mu\mathord{\pm}\mskip 2.0mu5.4$	$0.906\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.019$	$80.14\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.92$
Progressive	$1.091\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.014$	$1.386\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.014$	$28.22\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.20$	$74.4\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.7$	$0.912\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.005$	$78.53\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.71$
Emotion-first	$0.974\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.028$	$1.254\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.026$	$26.41\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.16$	$58.5\mskip 2.0mu\mathord{\pm}\mskip 2.0mu1.7$	$0.893\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.014$	$76.91\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.88$

Table 1: Formal FLUX comparison, including TIFA. Values are means

\pm

sample SD where repeated results are available.

Figure 4: Affect–semantic operating points on FLUX. The horizontal axis is the mean of V-Error and A-Error in V–A coordinate units; the vertical axis is pairwise DINO cosine similarity multiplied by 100. Lower horizontal values and higher vertical values are preferred. Vertical bars show the sample SD of DINO across the corresponding training or sampling repeats; EmotiCrafter is a single fixed-checkpoint result.

ID	SDE transitions	Optimized transitions	Transition 24
A	0–24	0–23	SDE
B	0–23	0–23	ODE
C	0–23	0–3	ODE
D	0–23	20–23	ODE
E	0–3	0–3	ODE
F	20–23	20–23	ODE

Table 2: Controlled sampling and optimization configurations.

ID	V-Error $\downarrow$	A-Error $\downarrow$	CLIPScore $\uparrow$	DINO $\uparrow$	CLIP-IQA $\uparrow$
A	$0.947\pm 0.046$	$1.191\pm 0.037$	$27.34\pm 0.19$	$64.7\pm 6.2$	$0.907\pm 0.021$
B	$0.963\pm 0.041$	$1.208\pm 0.034$	$27.58\pm 0.21$	$65.5\pm 5.4$	$0.906\pm 0.019$
C	$0.992\pm 0.058$	$1.251\pm 0.047$	$27.08\pm 0.24$	$61.6\pm 3.1$	$0.899\pm 0.017$
D	$1.024\pm 0.052$	$1.294\pm 0.043$	$28.12\pm 0.23$	$72.3\pm 2.4$	$0.913\pm 0.016$
E	$0.974\pm 0.028$	$1.254\pm 0.026$	$26.41\pm 0.16$	$58.5\pm 1.7$	$0.893\pm 0.014$
F	$1.052\pm 0.044$	$1.336\pm 0.038$	$29.51\pm 0.35$	$79.6\pm 2.1$	$0.917\pm 0.020$

Table 3: Effects of sampling and optimization windows on FLUX. Means

\pm

sample SD over three training seeds.

Method	Qwen V-MAE $\downarrow$	Qwen A-MAE $\downarrow$
EmotiCrafter	$1.541$	$1.685$
FLUX+VA-EIT (no RL)	$1.463\pm 0.054$	$1.602\pm 0.071$
Flow-GRPO A	$1.412\pm 0.082$	$1.637\pm 0.096$
Full-path B	$1.395\pm 0.059$	$1.571\pm 0.043$
Progressive	$1.367\pm 0.047$	$1.548\pm 0.068$
Emotion-first E	$1.329\pm 0.062$	$1.506\pm 0.051$

Table 4: Independent image-only affect evaluation with Qwen2.5-VL-7B-Instruct.

Method	Identity	Count	Action	Spatial	Drift $\downarrow$
EmotiCrafter	72.0	67.7	65.8	65.7	28.0
FLUX+VA-EIT (no RL)	83.3	83.8	81.7	81.0	16.7
Flow-GRPO	73.3	70.8	70.8	70.5	26.7
Progressive	80.0	79.2	79.2	76.2	20.0
Emotion-first	70.0	66.9	63.3	63.8	30.0
Semantic-first (ours)	97.3	97.7	95.8	95.2	2.7

Table 5: Expanded semantic audit: accuracy by prompt requirement and subject-drift rate (%).

Figure 5: Five-target evaluation at the four corners and center of the V–A grid. Panels report (a) principal-subject retention in percent, (b) CLIPScore (cosine similarity

\times 100

), and (c) DINO consistency (cosine similarity

\times 100

). Markers show means and horizontal error bars show sample SD across three seeds. Each method and target includes 132 images per seed (396 total).

Panel of Figure 6 — Figure 6: Paired scheduling comparison at three V–A targets. Each pair shares the prompt, target, initial noise, and sampling parameters. Emotion-first changes subject or composition; Semantic-first preserves the mountain landscape, astronaut, and boy–book–tree relations.

Width (steps)	V-Error $\downarrow$	A-Error $\downarrow$	CLIP $\uparrow$	DINO $\uparrow$	IQA $\uparrow$
W2 (22–23)	$1.097\pm 0.062$	$1.371\pm 0.051$	$\mathbf{29.84\pm 0.47}$	$\mathbf{80.9\pm 2.6}$	$\mathbf{0.921\pm 0.013}$
W4 (20–23)	$1.052\pm 0.044$	$1.336\pm 0.038$	$29.51\pm 0.35$	$79.6\pm 2.1$	$0.917\pm 0.020$
W6 (18–23)	$\mathbf{1.013\pm 0.059}$	$\mathbf{1.289\pm 0.047}$	$28.63\pm 0.51$	$77.8\pm 2.4$	$0.910\pm 0.018$
W8 (16–23)	$1.124\pm 0.073$	$1.409\pm 0.066$	$27.91\pm 0.68$	$73.4\pm 3.2$	$0.898\pm 0.031$

Table 6: Late-window width sensitivity on FLUX.

Method	V-Error $\downarrow$	A-Error $\downarrow$	CLIP $\uparrow$	DINO $\uparrow$	IQA $\uparrow$
M0: no RL	$1.274\pm 0.058$	$1.651\pm 0.043$	$29.07\pm 0.16$	$\mathbf{81.9\pm 0.6}$	$\mathbf{0.925\pm 0.007}$
M1: Full-path	$1.129\pm 0.091$	$1.452\pm 0.077$	$28.42\pm 0.25$	$75.2\pm 2.1$	$0.909\pm 0.019$
M2: Early	$1.092\pm 0.047$	$1.318\pm 0.039$	$27.33\pm 0.19$	$71.8\pm 1.4$	$0.916\pm 0.012$
M3: Late (ours)	$\mathbf{1.067\pm 0.031}$	$\mathbf{1.295\pm 0.028}$	$\mathbf{29.71\pm 0.21}$	$79.4\pm 1.2$	$0.920\pm 0.009$

Table 7: Second-backbone evaluation on SD3.5-Medium.

Figure 8: Measured training time and total GPU cost on eight A100 GPUs. All configurations use 300 formal iterations after 10 additional warm-up iterations. A/B use 24 policy-loss transitions per trajectory; C–F use four. Yellow identifies F.

Method	Affect match	Semantic preservation	Overall preference
EmotiCrafter	$3.21\pm 0.86$	$3.76\pm 0.74$	$3.41\pm 0.79$
FLUX+VA-EIT (no RL)	$3.67\pm 0.71$	$3.95\pm 0.63$	$3.72\pm 0.68$
Flow-GRPO	$\mathbf{4.22\pm 0.64}$	$3.02\pm 0.91$	$3.19\pm 0.85$
Progressive	$4.01\pm 0.78$	$3.61\pm 0.76$	$3.78\pm 0.74$
Emotion-first	$3.88\pm 0.69$	$3.33\pm 0.82$	$3.47\pm 0.77$
Semantic-first (ours)	$4.13\pm 0.58$	$\mathbf{4.29\pm 0.55}$	$\mathbf{4.24\pm 0.57}$

Table 8: Expanded human evaluation: 30 participants, 900 images, and three ratings per image.

Figure 9: Expanded human evaluation on a 1–5 scale. Thirty participants rate 900 images, with three ratings per image. Points show the mean of participant-level method means; error bars show their sample SD. Yellow identifies Semantic-first.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Comparison	CLIPScore	DINO	V-Error	A-Error
F $-$ B	$+1.93\;[1.68,\,2.18]$	$+14.1\;[12.4,\,15.8]$	$+0.089\;[0.048,\,0.130]$	$+0.128\;[0.085,\,0.171]$
F $-$ E	$+3.10\;[2.78,\,3.42]$	$+21.1\;[19.7,\,22.5]$	$+0.078\;[0.043,\,0.113]$	$+0.082\;[0.048,\,0.116]$
D $-$ C	$+1.04\;[0.76,\,1.32]$	$+10.7\;[9.4,\,12.0]$	$+0.032\;[0.012,\,0.052]$	$+0.043\;[0.022,\,0.064]$
F $-$ D	$+1.39\;[1.09,\,1.69]$	$+7.3\;[6.2,\,8.4]$	$+0.028\;[0.011,\,0.045]$	$+0.042\;[0.021,\,0.063]$
S2 $-$ S1	$+1.17\;[0.91,\,1.43]$	$+7.6\;[6.4,\,8.8]$	$-0.028\;[-0.058,\,0.002]$	$-0.045\;[-0.072,\,-0.018]$
B $-$ A	$+0.24\;[0.03,\,0.45]$	$+0.8\;[-0.3,\,1.9]$	$+0.016\;[-0.008,\,0.040]$	$+0.017\;[-0.007,\,0.041]$

Appendix

Table A.1: Paired FLUX differences and 95% prompt-bootstrap confidence intervals.

Baseline	Qwen V-MAE difference	Qwen A-MAE difference
EmotiCrafter	$-0.257\;[-0.312,\,-0.191]$	$-0.206\;[-0.274,\,-0.141]$
FLUX+VA-EIT (no RL)	$-0.179\;[-0.221,\,-0.132]$	$-0.123\;[-0.164,\,-0.087]$
Flow-GRPO A	$-0.128\;[-0.171,\,-0.082]$	$-0.158\;[-0.219,\,-0.102]$
Full-path B	$-0.111\;[-0.149,\,-0.077]$	$-0.092\;[-0.134,\,-0.055]$
Progressive	$-0.083\;[-0.124,\,-0.041]$	$-0.069\;[-0.116,\,-0.018]$
Emotion-first E	$-0.045\;[-0.083,\,-0.009]$	$-0.027\;[-0.068,\,0.019]$

Appendix

Table A.2: Independent affect evaluation: Ours minus each baseline, with 95% prompt-bootstrap confidence intervals.

Baseline	Difference	95% CI
EmotiCrafter	$+7.52$	$[7.18,\,7.89]$
FLUX+VA-EIT (no RL)	$+1.85$	$[1.49,\,2.18]$
Flow-GRPO A	$+2.59$	$[2.21,\,2.93]$
Full-path B	$+2.20$	$[1.86,\,2.57]$
Progressive	$+3.81$	$[3.47,\,4.16]$
Emotion-first E	$+5.43$	$[5.09,\,5.81]$

Appendix

Table A.3: TIFA paired differences: Semantic-first minus each baseline.

Baseline	Affect match	Semantic preservation	Overall preference
EmotiCrafter	$+0.92\;[0.61,\,1.23]$	$+0.53\;[0.31,\,0.75]$	$+0.83\;[0.55,\,1.11]$
FLUX+VA-EIT (no RL)	$+0.46\;[0.22,\,0.70]$	$+0.34\;[0.15,\,0.53]$	$+0.52\;[0.30,\,0.74]$
Flow-GRPO	$-0.09\;[-0.31,\,0.13]$	$+1.27\;[0.98,\,1.56]$	$+1.05\;[0.77,\,1.33]$
Progressive	$+0.12\;[-0.08,\,0.32]$	$+0.68\;[0.45,\,0.91]$	$+0.46\;[0.25,\,0.67]$
Emotion-first	$+0.25\;[0.03,\,0.47]$	$+0.96\;[0.70,\,1.22]$	$+0.77\;[0.52,\,1.02]$

Appendix

Table A.4: Human evaluation: Ours minus each baseline, with paired 95% confidence intervals.

Metric	M3 $-$ M1	M3 $-$ M2
V-Error	$-0.062\;[-0.098,\,-0.021]$	$-0.025\;[-0.049,\,-0.002]$
A-Error	$-0.157\;[-0.204,\,-0.113]$	$-0.023\;[-0.046,\,0.003]$
CLIPScore	$+1.29\;[1.01,\,1.58]$	$+2.38\;[2.07,\,2.69]$
DINO	$+4.2\;[3.1,\,5.4]$	$+7.6\;[6.7,\,8.5]$
Qwen V-MAE	$-0.103\;[-0.142,\,-0.062]$	$-0.027\;[-0.051,\,-0.004]$
Qwen A-MAE	$-0.092\;[-0.131,\,-0.048]$	$-0.017\;[-0.039,\,0.008]$

Appendix

Table A.5: SD3.5-Medium: paired differences with 95% prompt-bootstrap confidence intervals.

Metric	W2 $-$ W4	W6 $-$ W4	W8 $-$ W4
V-Error	$+0.045\;[0.011,\,0.083]$	$-0.039\;[-0.075,\,-0.005]$	$+0.072\;[0.029,\,0.111]$
A-Error	$+0.035\;[0.007,\,0.069]$	$-0.047\;[-0.082,\,-0.013]$	$+0.073\;[0.032,\,0.119]$
CLIPScore	$+0.33\;[0.12,\,0.54]$	$-0.88\;[-1.14,\,-0.63]$	$-1.60\;[-1.93,\,-1.28]$
DINO	$+1.3\;[0.5,\,2.2]$	$-1.8\;[-2.7,\,-1.1]$	$-6.2\;[-7.8,\,-4.7]$
Qwen V-MAE	$+0.034\;[-0.004,\,0.072]$	$-0.028\;[-0.059,\,0.003]$	$+0.041\;[0.008,\,0.077]$
Qwen A-MAE	$+0.018\;[-0.011,\,0.053]$	$-0.038\;[-0.069,\,-0.006]$	$+0.033\;[-0.005,\,0.069]$

Appendix

Table A.6: Width sensitivity: paired differences and 95% prompt-bootstrap intervals.

ID	SDE steps	V-Error $\downarrow$	A-Error $\downarrow$	CLIP $\uparrow$	DINO $\uparrow$	IQA $\uparrow$
S0	None (ODE)	$1.402\pm 0.048$	$1.710\pm 0.036$	$29.27\pm 0.27$	$81.6\pm 0.9$	$0.924\pm 0.024$
S1	0–3	$1.439\pm 0.055$	$1.748\pm 0.044$	$27.91\pm 0.25$	$72.8\pm 2.1$	$0.909\pm 0.018$
S2	20–23	$1.411\pm 0.043$	$1.703\pm 0.034$	$29.08\pm 0.24$	$80.4\pm 1.5$	$0.921\pm 0.017$

Appendix

Table B.1: Sampling-window intervention with a fixed no-RL FLUX+VA-EIT model.

Inference	V-Error $\downarrow$	A-Error $\downarrow$	CLIPScore $\uparrow$	DINO $\uparrow$	CLIP-IQA $\uparrow$
F-all	$1.052\pm 0.044$	$1.336\pm 0.038$	$29.51\pm 0.35$	$79.6\pm 2.1$	$0.917\pm 0.020$
F-window	$1.058\pm 0.045$	$1.342\pm 0.039$	$29.24\pm 0.33$	$79.0\pm 2.0$	$0.915\pm 0.019$

Appendix

Table B.2: Inference-time use of the trained F weights.

Width	Optimized steps	s/iteration	Wall-clock h	Total GPU-h
W2	2	13.85	1.15	9.23
W4	4	18.76	1.56	12.51
W6	6	24.32	2.03	16.21
W8	8	29.68	2.47	19.79

Appendix

Table B.3: Training cost across window widths on

8\times

A100.

Window	V-Error $\downarrow$	A-Error $\downarrow$	CLIP $\uparrow$	TIFA $\uparrow$	IQA $\uparrow$	DINO $\uparrow$
Early	$\mathbf{0.978}\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.025$	$\mathbf{1.262}\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.032$	$25.641\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.280$	$60.80\mskip 2.0mu\mathord{\pm}\mskip 2.0mu1.00$	$0.885\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.008$	$55.23\mskip 2.0mu\mathord{\pm}\mskip 2.0mu2.10$
Middle	$1.138\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.027$	$1.485\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.046$	$27.725\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.223$	$69.40\mskip 2.0mu\mathord{\pm}\mskip 2.0mu1.20$	$0.903\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.011$	$70.97\mskip 2.0mu\mathord{\pm}\mskip 2.0mu2.75$
Late	$1.296\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.021$	$1.582\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.039$	$\mathbf{28.936}\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.256$	$\mathbf{75.20}\mskip 2.0mu\mathord{\pm}\mskip 2.0mu1.40$	$\mathbf{0.916}\mskip 2.0mu\mathord{\pm}\mskip 2.0mu0.004$	$\mathbf{78.27}\mskip 2.0mu\mathord{\pm}\mskip 2.0mu3.15$

Appendix

Table D.1: Three-seed window-position pilot (200 updates per run, fixed width

w=4

Method	Target $(V,A)$	V-Error $\downarrow$	A-Error $\downarrow$	CLIP $\uparrow$	DINO $\uparrow$	Retention (%) $\uparrow$
Emotion-first	$(3,3)$	0.918	1.371	$23.914\pm 0.34$	$49.1\pm 5.2$	$68.9\pm 4.2$
Emotion-first	$(-3,-3)$	1.003	0.982	$24.408\pm 0.28$	$50.3\pm 4.1$	$73.0\pm 3.4$
Emotion-first	$(3,-3)$	0.884	1.268	$24.739\pm 0.41$	$51.6\pm 5.8$	$73.7\pm 5.7$
Emotion-first	$(-3,3)$	1.025	1.105	$23.885\pm 0.37$	$49.7\pm 4.5$	$71.2\pm 3.8$
Emotion-first	$(0,0)$	0.822	1.082	$27.243\pm 0.22$	$58.7\pm 3.3$	$77.3\pm 3.0$
Semantic-first	$(3,3)$	0.996	1.381	$28.831\pm 0.31$	$76.4\pm 2.9$	$94.7\pm 2.7$

Appendix

Table D.2: Three-seed evaluation at extreme and neutral V–A targets.

Width	Qwen V-MAE $\downarrow$	Qwen A-MAE $\downarrow$
W2	$1.318\pm 0.058$	$1.497\pm 0.049$
W4	$1.284\pm 0.036$	$1.479\pm 0.039$
W6	$\mathbf{1.256\pm 0.042}$	$\mathbf{1.441\pm 0.055}$
W8	$1.325\pm 0.061$	$1.512\pm 0.057$

Appendix

Table D.3: Independent affect evaluation for the width study.

Method	Qwen V-MAE $\downarrow$	Qwen A-MAE $\downarrow$
M0: no RL	$1.582\pm 0.069$	$1.613\pm 0.051$
M1: Full-path	$1.427\pm 0.084$	$1.544\pm 0.093$
M2: Early	$1.351\pm 0.052$	$1.469\pm 0.068$
M3: Late (ours)	$\mathbf{1.324\pm 0.041}$	$\mathbf{1.452\pm 0.047}$

Appendix

Table D.4: Independent affect evaluation on SD3.5-Medium.

ID	Optimized steps	s/iteration	Wall-clock h	Total GPU-h	Time saved
A	24	31.84	2.65	21.23	–
B	24	31.27	2.61	20.85	1.8%
C	4	24.96	2.08	16.64	21.6%
D	4	24.51	2.04	16.34	23.0%
E	4	19.37	1.61	12.91	39.2%
F	4	18.76	1.56	12.51	41.1%

Appendix

Table D.5: Training cost on

8\times

NVIDIA A100 for 300 formal outer iterations.

Controllable Image Generation

Controllable image generation steers images using spatial, visual, semantic, stylistic, or physical constraints beyond ordinary text prompts.

See all 215 papers

Balancing Emotional Alignment and Semantic Consistency in Image Generation via Reinforcement Learning with Valence-Arousal Anchoring
Sep 11, 2026Jisheng Dang, Zhenxuan Wang, Bin Li +3Conditional Image Generation Image Generation
School of Information Science & Engineering, Lanzhou University, Lanzhou 730000, China · School of Environmental and Spatial Informatics, China University of Mining and Technology, Xuzhou 221116, China · College of Computer Science and Software Engineering, Shenzhen University, Shenzhen 518060, China +2
$λ$ -Controlled GRPO: Turning Flow-Matching Ratio Instability into a Budgeted Resource
Sep 18, 2026Yufeng Wang, Parivesh Priye, Meeshawn Marathe +1Flow Matching T2I Generation
Stony Brook University · Georgia Institute of Technology · Rivian and Volkswagen Group Technologies
MixGRPO: Unlocking Flow-based GRPO Efficiency with Mixed ODE-SDE
Jul 29, 2025Junzhe Li, Yutao Cui, Tao Huang +8Diffusion Model Alignment Flow Matching
Hunyuan, Tencent, China · School of Computer Science, Peking University, China · Computer Center, Peking University, China +2

Sep 11, 2026J/K move · Enter open · S save

Balancing Emotional Alignment and Semantic Consistency in Image Generation via Reinforcement Learning with Valence-Arousal Anchoring

Jisheng Dang, Zhenxuan Wang, Bin Li, Ronghao Lin+2

School of Information Science & Engineering, Lanzhou University, Lanzhou 730000, China · School of Environmental and Spatial Informatics, China University of Mining and Technology, Xuzhou 221116, China · College of Computer Science and Software Engineering, Shenzhen University, Shenzhen 518060, China+2

Conditional Image Generation Image Generation RL for Image Generation Controllable Image Generation

Open paper AskPDF arXiv

Continuous emotion control in text-to-image generation requires a model to improve affective alignment without changing the objects, layout, or scene described by the prompt. Existing supervised emotion-injection methods often optimize feature-space proxies and may therefore exhibit emotion-semantic drift, in which stronger emotional conditioning is accompanied by unintended content changes. We address this problem with a flow-matching image-generation framework that combines continuous valence-arousal (VA) conditioning, Group Relative Policy Optimization (GRPO), and a neutral semantic anchor. The deterministic probability-flow ODE is converted into a marginal-preserving SDE, yielding non-degenerate transition densities for trajectory sampling and policy-ratio estimation. A frozen CLIP-based VA regressor supplies a terminal reward measuring the distance between the predicted and target VA coordinates, while an image generated from the same prompt under zero VA conditioning provides a feature-space reference for semantic preservation. A reduced denoising schedule is used for online RL sampling, whereas the original schedule is retained at inference. Experiments on 3,300 prompt-emotion combinations show substantially lower valence and arousal errors than the VA-conditioned baseline and an improved CLIPScore relative to EmotiCrafter, with a measurable trade-off in reference-free image quality. The results support anchor-regularized Flow-GRPO as a practical approach to balancing emotional alignment and semantic consistency in continuous-affect image synthesis.

Abstract

Controllable Image Generation

Explore similar work

Balancing Emotional Alignment and Semantic Consistency in Image Generation via Reinforcement Learning with Valence-Arousal Anchoring

λλλ-Controlled GRPO: Turning Flow-Matching Ratio Instability into a Budgeted Resource

MixGRPO: Unlocking Flow-based GRPO Efficiency with Mixed ODE-SDE

$λ$ -Controlled GRPO: Turning Flow-Matching Ratio Instability into a Budgeted Resource