WorksComputer Science
Mem2Fair: Online Fairness Control for Text-to-Video Generation via Fair Memory
Abstract
Text-to-video (T2V) generation models can produce systematically unfair demographic outcomes when prompts leave protected attributes unspecified. For example, prompts describing nurses without specifying gender may disproportionately generate videos depicting female nurses. Mitigating unfairness in T2V generation is particularly challenging: training-based methods that modify model internals require architecture-specific access and risk disrupting temporal consistency, while sampling-based methods often incur costly pre-deployment video generation. We present the first conditional online approachability formulation of attribute-level fairness for T2V generation. Building on this formulation, we introduce Mem2Fair, a plug-and-play framework that maintains concept-conditioned Fair Memory. Using generated videos at inference time as feedback, Mem2Fair selects prompt-level attribute interventions based on the deviation from each concept's prescribed fairness target, without training or separate pre-deployment sampling. Experiments demonstrate that Mem2Fair consistently improves fairness over the unmodified generator and existing baselines across multiple demographic attributes, prompt settings, and target distributions, while maintaining comparable prompt-video alignment and video quality. We further demonstrate its effectiveness on commercial black-box T2V models.
Introduction
Text-to-video (T2V) models have developed rapidly in recent years, as illustrated by recent large-scale systems such as Wan 2.2, Seedance 2.0 and UniVideo [25, 32, 33]. As these models become increasingly capable and widely deployed, ensuring fairness in their generated content has become an important concern. Biases embedded in large-scale training data and complex model architectures can be reflected in generated content, leading attribute-unspecified prompts to produce systematically imbalanced outcomes across different attribute groups. Videos generated from prompts about specific concepts may exhibit imbalanced representation across gender, age, and race [20, 36].
These fairness concerns are well documented in text-to-image (T2I) models, where a growing body of work has proposed mitigation methods [11, 15, 18, 19]. However, T2V fairness cannot be treated as a direct extension of T2I fairness because it presents three distinct challenges: (i) Temporal fidelity. Interventions that directly alter text embeddings, latent representations, or denoising trajectories may also affect temporal properties of generated videos, including inter-frame identity, motion, and consistency [14]. (ii) Dependence on model internals. Fine-tuning parameters, editing text encoders or attention modules, and changing denoising dynamics require access to architecture-specific components [10, 15, 19, 28, 31], limiting their applicability to advanced closed-source T2V systems. (iii) Pre-deployment sampling cost. Many fairness methods rely on generated samples to train or construct their interventions. For T2V systems, producing these samples incurs substantial additional generation cost before deployment.
To address these challenges, we formulate fair T2V generation as a conditional online approachability problem [3, 8]. This formulation treats fairness as an outcome-feedback problem in which generated videos provide observations that guide subsequent interventions for the corresponding semantic concept. By leveraging feedback collected during the sequential generation process, fairness control does not require exhaustive offline sampling across different prompts, events, and interventions, thereby avoiding the associated upfront sampling burden.
Building on this formulation, we propose Mem2Fair, a plug-and-play closed-loop framework that interacts with T2V generators only through input prompts and generated videos. It maintains a separate Fair Memory for each semantic concept, tracking realized attribute outcomes. At each generation round, an online approachability-based controller uses the discrepancy between the concept’s empirical outcome distribution and the prescribed target to select an attribute value, which is then incorporated into the input prompt through prompt optimization. After generation, an external attribute classifier observes the outcome and updates the corresponding memory, providing feedback for subsequent generations. This input/output-only design requires no access to T2V model internals, leaves the generator’s temporal process unchanged, and supports both open- and closed-source systems.

Figure 1. The nurse example illustrates demographic bias from a gender-unspecified prompt. Existing interventions face model-access constraints, fidelity disruption, or costly pre-sampling. Mem2Fair uses Fair Memory and outcome feedback to guide prompt optimization, steering demographic proportions toward a prescribed target across requests.
Our main contributions are:
To the best of our knowledge, we are the first to formulate attribute-level fairness in T2V generation as a conditional online approachability problem, using realized video outcomes as feedback to guide sequential decisions.
We develop Mem2Fair, a plug-and-play framework that uses concept-conditioned Fair Memory to guide attribute selection and optimizes the input prompt to incorporate the selected attribute value, without modifying the video generator. We further establish a finite-sample bound on the expected concept-level fairness discrepancy.
We evaluate Mem2Fair across multiple T2V generators, protected attributes, prompt settings, and target distributions, demonstrating consistent fairness improvements with comparable video quality and prompt–video alignment. We further validate its applicability to commercial black-box T2V models.
Related Work
Bias in Image and Video Generation and Mitigation Methods
Text-to-image (T2I) diffusion models systematically under-represent particular demographic groups and amplify social stereotypes, as quantified by auditing studies and fairness benchmarks [2, 6, 20, 27, 30]. Existing mitigation methods intervene at different points in the generation pipeline. Model-adaptation methods fine-tune diffusion parameters or learn inclusive prompt tokens [28, 35]; conditioning-space methods edit text embeddings or encoders [11, 12, 15, 19]; and inference-time methods modify guidance, attention, latent directions, or sampling [7, 10, 18, 22, 31]. Some strategies can theoretically be adapted to T2V, but direct transfer may inherit challenges, including temporal-fidelity loss, dependence on model-specific internals, and high pre-deployment sampling costs when training or calibration is required.
Existing research on T2V fairness encompasses both fairness assessment and mitigation. Fairness assessment studies measure attribute-level biases in generated videos, including gender and racial associations [21, 29], while VMDT [23] evaluates fairness as part of broader model trustworthiness. At the training stage, controllable preference modeling has been explored by constructing gender-controlled preference data and training reward models to guide video-model alignment [4]. However, this strategy requires curated preference data and additional reward-model and video-model training. At inference time, explicit debiasing prompts can reduce measured attribute associations but may exacerbate them for certain occupations [29]. FairT2V [36] instead neutralizes prompt embeddings and applies interventions at selected denoising steps, requiring access to model internals.
Memory and Prompt Optimization for Generation
Memory has supported several objectives in generation, including preserving long-range video context and reusing information from earlier prompts. Context-as-Memory [34] retrieves relevant past frames for scene-consistent long-video generation, while VideoMemory [37] stores and updates entity representations to maintain consistency across narrative shots. For compositional T2V generation, TTOM [24] stores test-time optimization parameters under prompt-derived layout keys and reuses them to improve text-video alignment.
Prompt optimization is a practical way to improve conditional generation by rewriting or expanding user prompts to better match a generator’s learned prompt distribution and to specify relevant semantic details. For instance, Promptist [16] optimizes prompts toward those preferred by human users, aiming to improve text-to-image generation quality, while RAPO [13] adapts prompts to the text-to-video model’s prompt distribution to improve the static and dynamic quality of generated videos.
Online Fairness and Blackwell Approachability. At deployment, T2V systems receive a stream of prompts. Fairness concerns the distribution of realized demographic outcomes across repeated generations, while each video provides only one observation. This connects the problem to online fairness, which studies cumulative outcome constraints [8], and fairness over sequences of decisions [1, 5]. Blackwell approachability provides a geometric framework for steering a running vector average toward a target set [3]. These frameworks offer tools for analyzing distributional fairness under sequential feedback.
Preliminaries
Fairness objective. Let the protected attribute take values in , and let be the -dimensional probability simplex. The prescribed target distribution is . For each semantic concept , let be the distribution of detected outcomes over attribute-unspecified generations. The ideal fairness objective is
Quality objective. Let and be the controlled and original videos for the same prompt , and let measure video quality. Ideally, fairness interventions should preserve video quality within a small tolerance:
Blackwell approachability. Fix a concept and let index its eligible generations. Write for the one-hot encoding of the observed outcome ; its running average is the empirical outcome distribution. Blackwell approachability seeks actions for which this average approaches a closed target set [3]:
For fairness, is a target region around the prescribed distribution .
Method
As illustrated in Figure 2, Mem2Fair performs fairness-aware prompt optimization using concept-conditioned Fair Memory. The memory guides attribute selection, and an LLM prompt optimizer incorporates the selected value into the input prompt while preserving its concept and context. After generation, the detected attribute updates the corresponding memory, closing the loop for subsequent requests.
Figure 2. Overview of Mem2Fair. Concept-conditioned Fair Memory guides fairness-aware prompt optimization through attribute selection and prompt optimization. Detected outcomes update the corresponding memory for subsequent requests.
Concept-Conditioned Fair Memory
At request , the prompt maps to a concept . We maintain a separate state for each concept because pooling outcomes can let frequent concepts mask bias in rarer ones. We only intervene on prompts that leave the protected attribute unspecified, while prompts that explicitly specify an attribute value are left unchanged.
An outcome is valid if the detector returns an attribute label; cases with no detected face or failed classification are excluded from Fair Memory updates. Let index the attribute-unspecified requests for concept with valid detected outcomes up to time , and let . For detected outcome , write for its one-hot encoding, with . The empirical outcome distribution is
As a practical control target, we relax Equation (1) to an acceptable fairness target set for each concept , assumed nonempty, closed, and convex, and to contain the prescribed target distribution . Rather than requiring exact equality with , the controller seeks to steer the empirical outcome distribution toward . The Euclidean projection onto defines the correction residual:
The residual vanishes on , and its norm measures the fairness gap. Fair Memory stores , providing the state for subsequent requests of the same concept. The target set is fixed per concept and supplied to the controller as an input alongside .
Fairness-Aware Prompt Optimization
For an attribute-unspecified prompt , fairness-aware prompt optimization translates the concept-specific fairness residual into an explicit attribute intervention in the input text.
Fairness-guided attribute selection. For a concept with existing Fair Memory, the controller selects the attribute that minimizes the fairness residual score:
This rule favors attributes that are underrepresented relative to the projected fairness target. When multiple attributes attain the same minimum score, ties are broken by sampling according to their relative probabilities in the target distribution . If the empirical distribution is already within , then , all attributes have equal scores, and the controller samples from . For a new concept (), the controller samples before applying the same optimizing procedure.
For example, use male/female coordinates with and . Suppose . Since the male proportion exceeds the upper bound 0.6, the closest acceptable distribution is . Equation (6) then gives the residual scores . Equation (7) therefore selects female, whose negative residual indicates underrepresentation relative to the projected target.
Semantics-preserving prompt optimization. The selected attribute is incorporated into the original prompt by an LLM optimizer:
Using a fixed instruction template, the LLM optimizer is instructed to make the selected attribute explicit for the target person while preserving the original concept, action, and scene context. For example, selecting female can turn “Generate a video of a doctor rushing through a hospital corridor.” into “Generate a video of a female doctor rushing through a hospital corridor.” The optimized prompt is then passed to the video generator.
Outcome feedback and memory update. The generator produces , and an external classifier observes . If the detection is invalid, the video is returned without updating Fair Memory. Otherwise, since the generator may not realize the selected attribute, memory is updated with the detected outcome : increment , recompute using Equation (5), and update the residual. Other concepts’ states remain unchanged. This feedback uses the requested video itself, requiring no separate calibration generations. Algorithm 1 summarizes the procedure, updating Fair Memory in place.
Algorithm 1 Mem2Fair: one request t
Require: prompt $x_t$; Fair Memory $\mathcal{M}$; target $\pi$; target sets $\{\mathcal{F}_c\}_c$
Ensure: video $v_t$ and updated Fair Memory $\mathcal{M}$
1: if $x_t$ specifies the protected attribute then
2: return $(G(x_t),\mathcal{M})$
3: end if
4: $c_t\leftarrow C(x_t)$
5: if $c_t$ has no valid observations in $\mathcal{M}$ then
6: sample $a_t\sim\pi$
7: else
8: choose $a_t\in\arg\min_{a\in A}\langle g^{c_t},h(a)\rangle$
9: end if
10: $v_t\leftarrow G(\mathcal{R}(x_t,a_t));\ Y_t\leftarrow D(v_t)$
11: if $Y_t\in A$ then
12: update $N^{c_t}$ and $\hat{p}^{c_t}$ with $h(Y_t)$
13: recompute $g^{c_t}$ using {xref}`Equation 6`
14: store $\mathcal{M}(c_t)\leftarrow(N^{c_t},\hat{p}^{c_t},g^{c_t})$
15: end if
16: return $(v_t,\mathcal{M})$
Conditional Fairness Guarantee
Fix a concept and index its valid detected outcomes by . Let , and denote the corresponding empirical distribution, projection, and residual by , , and , respectively. Let denote the history available before the -th decision, including the current prompt.
The first valid observation initializes Fair Memory, so the directional condition is imposed from the second valid observation onward.
Assumption 1 (Approximate directional condition). For every concept and local round ,
where is the directional slack.
Theorem 1 (Approximate conditional approachability). Under Assumption 1, after valid detected outcomes,
where is valid on the probability simplex.
The bound separates directional slack from finite-sample error, yielding an expected distance of at most . Under perfect attribute compliance, the projection property and ensure that a corrective direction always exists, giving and an convergence rate. Imperfect compliance may introduce nonzero slack, leaving an asymptotic upper bound of on the expected distance.
The guarantee applies to each concept’s detected-outcome distribution, including the cold-start observation. The proof and the relationship between directional slack and attribute selection are provided in Appendix § 8.
Experiments
Experimental Setup
Model and data. We use Wan2.2-5B as the base text-to-video generator [32]. Our occupation set is drawn from the Stable Bias dataset [20]. We evaluate three protected-attribute tasks: gender ({female, male}), race ({white, black, asian, indian, middle eastern, latino hispanic}), and age ({young, middle-aged, old}).
Compared methods Vanilla uses the unmodified Wan2.2 model as the reference. We apply FairGen [18], FairImagen [11], and LightFair [15] to the same Wan2.2 backbone for comparison.
Metrics. Let be an evaluation prompt set, let be the number of videos generated per prompt, let be the number with a valid attribute detection for prompt , and let be the detected attribute of valid video for prompt . For a prescribed target distribution , define
The Normalized Parity (NP) pools all generated videos and measures the deviation from , following the normalized-discrepancy formulation of Teo et al. [30] and its uniform-target instantiation in FairImagen [11], which we generalize from a uniform reference to an arbitrary target :
Here denotes an exact match to the target distribution, and larger values indicate greater global fairness. Detailed derivation in Appendix § 9. A prompt with contributes no attribute information, so we exclude it from per-prompt statistics; the Bias Score therefore averages over the prompts with at least one valid detection:
Lower Bias Score indicates that individual prompts produce distributions closer to . We also report as prompt–video semantic alignment, computed by comparing the prompt representation with the visual representation of the video’s middle frame:
Finally, we report the VBench [17] dimensions subject_consistency, motion_smoothness, and imaging_quality, for which higher values are better.
Implementation Details. We use Qwen3-0.6B to map each prompt to an occupation concept by matching it against a predefined occupation vocabulary, and to rewrite prompts using a fixed instruction template. DeepFace [26] provides the output-side attribute classifier, with RetinFace [9] as the face detection backend. Wan2.2 generates videos with 49 frames at a resolution of 1280 704. Unless otherwise specified, all settings use one test prompt per occupation and eight independent random seeds per prompt, with the same prompts and seed protocol across methods. Fair Memory is initialized from scratch for each experimental setting. We compute CLIPsim using CLIP ViT-L/14. Across all attributes, concepts, and target distributions, we use a normalized TV tolerance :
with Euclidean projection onto .
Main Results
In-Distribution Gender Control. We evaluate gender control on 50 seen occupation concepts under the uniform target .
Table 1 compares Mem2Fair with the unmodified generator and baselines. Mem2Fair raises NP from 0.336 to 0.917 and lowers Bias Score from 0.384 to 0.052, a larger
| Method | Fairness | Alignment | Quality | |||||
| Vanilla | 0.336 | 0.384 | ||||||
| FairGen | 0.500 | |||||||
| FairImagen | 0.287 | |||||||
| LightFair | 0.342 | |||||||
| Mem2Fair |
Table 1. Gender results on the 50 in-distribution occupations under the uniform target distribution . Avg. is the arithmetic mean of the three quality metrics (Subj., Motion, and Imaging). Bold and _underlined_ means indicate the best and second-best results across all methods, respectively.
margin than any baseline achieves. Mem2Fair also exceeds all three fairness baselines in CLIPsim (0.1941) and average quality (0.8975).
FairGen incurs the largest quality loss, particularly in imaging quality. FairImagen also reduces CLIPsim and average video quality relative to Vanilla. Changing denoising or text embeddings may also affect visual details and motion consistency in Wan2.2. We also provide a qualitative video quality comparison in Figure 6.

Figure 6. Video quality comparison for cook. Each row shows one generated video; columns show frames in chronological order from left to right, with timestamps matched across methods.
LightFair preserves relatively high video quality but provides limited fairness correction in this setting. This may reflect a limitation of transferring its two-stage guidance scheme to Wan2.2, where restricting debiased conditioning to selected generation stages may provide insufficient control over demographic outcomes.
Multi-Attribute Control on Race and Age. We evaluate race () and age () on the same 50 seen occupation concepts with a uniform target for each attribute.
Figure 4 reports the results. Mem2Fair achieves the highest NP and lowest Bias Score for both attributes relative to the strongest baseline. Figure 3 illustrates early demographic representation across consecutive generations for the example concepts, complementing the aggregate distributional results.
Figure 3. Early demographic representation under uniform target distributions. Each group shows the first four consecutive generations for a concept and method, with one representative frame per video. Mem2Fair starts each stream with empty Fair Memory.
Figure 4. Race (, left) and age (, right) results on the 50 seen occupation concepts under uniform target distributions. Higher NP and lower Bias are better. Red stars denote Mem2Fair.
Generalization to Unseen Occupations. The Unseen Concept setting evaluates gender control on 25 held-out occupations under the uniform target.
In Table 2, Mem2Fair achieves the best NP = 0.693 and Bias Score = 0.173. The baselines show limited fairness transfer to unseen occupations, yielding even higher Bias Scores. This is likely because FairGen’s sampled attribute adapters and LightFair’s fixed two-stage guidance do not adjust to each occupation’s realized output history. Under concept shift, this may leave occupation-specific control errors uncorrected, allowing within-occupation imbalance to persist or worsen. Mem2Fair instead uses concept-specific outcome feedback to steer subsequent generations toward the target.
Table 2. Generalization and robustness results for gender under the uniform target distribution . Unseen Concept: 25 held-out occupations with simple prompts. Complex Prompt: the 50 occupations of the Main setting, queried with richer action and contextual descriptions. Avg. is the arithmetic mean of the three quality metrics (Subj., Motion, and Imaging). Bold and <u>underlined</u> means indicate the best and second-best results within each setting, respectively.
| Setting | Method | Fairness | Alignment | Quality | ||||
| Unseen Concept | Vanilla | 0.290 | 0.375 | 0.1919 ± 0.0355 | 0.9814 ± 0.0124 | 0.9926 ± 0.0042 | 0.7380 ± 0.0411 | 0.9040 |
| FairGen | 0.290 | 0.500 | 0.1892 ± 0.0347 | 0.9481 ± 0.0116 | 0.9801 ± 0.0021 | 0.6431 ± 0.0359 | 0.8571 | |
| FairImagen | 0.250 | 0.385 | 0.1839 ± 0.0378 | 0.9708 ± 0.0108 | 0.9721 ± 0.0047 | 0.7069 ± 0.0446 | 0.8833 | |
| LightFair | 0.200 | 0.500 | 0.1876 ± 0.0362 | <u>0.9811</u> ± 0.0089 | 0.9853 ± 0.0033 | <u>0.7174</u> ± 0.0358 | <u>0.8946</u> | |
| Mem2Fair | 0.693 | 0.173 | 0.1889 ± 0.0378 | 0.9746 ± 0.0122 | <u>0.9913</u> ± 0.0033 | 0.7100 ± 0.0407 | 0.8920 | |
| Complex Prompt | Vanilla | <u>0.319</u> | <u>0.388</u> | 0.2515 ± 0.0333 | 0.9738 ± 0.0202 | 0.9964 ± 0.0046 | 0.7368 ± 0.0419 | 0.9023 |
| FairGen | 0.267 | 0.500 | 0.2409 ± 0.0355 | 0.9502 ± 0.0211 | <u>0.9896</u> ± 0.0038 | 0.6019 ± 0.0421 | 0.8472 | |
| FairImagen | 0.306 | <u>0.362</u> | <u>0.2451</u> ± 0.0331 | 0.9631 ± 0.0208 | 0.9858 ± 0.0043 | 0.6756 ± 0.0458 | 0.8748 | |
| LightFair | 0.227 | 0.500 | 0.2316 ± 0.0338 | <u>0.9705</u> ± 0.0199 | 0.9661 ± 0.0037 | 0.7061 ± 0.0407 | 0.8809 | |
| Mem2Fair | 0.856 | 0.075 | 0.2427 ± 0.0405 | 0.9611 ± 0.0181 | 0.9877 ± 0.0035 | <u>0.7271</u> ± 0.0295 | <u>0.8920</u> |
Table 2.
Robustness to Complex Prompts. The Complex Prompt setting evaluates the same 50 concepts as the in-distribution setting under the uniform target, with prompts enriched by occupation-specific actions and contextual descriptions. Methods requiring training or calibration use corresponding complex prompts during preparation.
As reported in Table 2, Mem2Fair attains the strongest fairness results in this setting, with the highest average quality among baselines. FairImagen retains relatively strong prompt alignment but offers limited fairness correction. This may be because suppressing demographic directions in the embedding space does not directly enforce balanced demographic outcomes in the video generator.
Non-Uniform Target Control. We evaluate gender control on the same 50 seen occupation concepts with the non-uniform target . FairImagen does not support non-uniform targets and is omitted from this comparison.
Table 3 shows that Mem2Fair follows the prescribed target more closely than the supported baselines, reaching and Bias Score while preserving prompt-video alignment and imaging quality. This supports the controller’s ability to track prescribed demographic proportions beyond uniform parity, consistent with computing corrections relative to the target-dependent fairness set.
Table 3. Non-Uniform Target Control on the 50 seen occupation concepts under . Average scores are reported in the Quality group. FairImagen does not support non-uniform targets.
| Method | Fairness | Alignment | Quality | |
| Vanilla | 0.193 | 0.605 | 0.1991 | 0.8982 |
| FairGen | 0.193 | 0.678 | 0.1902 | 0.8310 |
| FairImagen | — | — | — | — |
| LightFair | 0.204 | 0.661 | 0.1873 | 0.8919 |
| Mem2Fair | 0.871 | 0.097 | 0.1914 | 0.8846 |
Table 3.
Black-Box Generator Deployment. We evaluate Mem2Fair on Seedance 2.02 and Kling v2.63 using the same 50 seen occupation concepts under the uniform target. The APIs are accessed only through inputs and outputs, without inspecting model parameters or internal representations.
Figure 5 (left) shows that Mem2Fair improves fairness on both APIs: on Seedance 2.0, NP rises from to and Bias Score falls from to ; on Kling v2.6, NP rises from to and Bias Score falls from to . The gains across APIs with different initial imbalances suggest that outcome feedback can support generator-specific correction using only input/output access.

Figure 5. Left: Black-box fairness on Seedance 2.0 and Kling v2.6 under the uniform gender target. Arrows connect Original outputs to outputs with Mem2Fair (red stars). Right: Ablation changes relative to the full method (NP = 0.917, Bias = 0.052). The zero line denotes the full method.
Ablation Study
We evaluate the full method and three ablated variants, each defined as follows. Full method uses a separate Fair Memory for each concept and updates it with the attribute detected from each generated video. w/o Outcome Feedback updates the memory with the requested attribute rather than the attribute detected from the generated video. w/o Concept Conditioning uses one shared memory across all concepts instead of maintaining concept-specific memories. w/o Memory & Detector directly samples an attribute from global target distribution and rewrites the prompt.
Figure 5 (right) reports metric changes relative to the full method, which achieves NP = and Bias Score = .
All ablated variants are worse than the full method. Sharing one memory across concepts still gives a moderate overall NP (0.820), but it increases the within-concept imbalance, producing the highest Bias Score among the ablated variants (0.239). Using the requested attribute as feedback gives NP = 0.793 and Bias = 0.112; when the generator realizes an attribute imperfectly, this variant records the request instead of the realized outcome and therefore cannot correct the resulting error. Removing both Memory and Detector gives the lowest ablation NP (0.764): direct sampling and rewriting ignores concept-specific history and realized-outcome feedback, leaving the generator’s demographic bias uncorrected. Together, these results show that concept-specific state and realized-outcome feedback each contribute to fairness control.
Conclusion
We formulate fair text-to-video generation as a conditional online approachability problem and introduce Mem2Fair, a plug-and-play framework that uses concept-conditioned Fair Memory for sequential fairness control. By operating solely through input prompts and generated videos, Mem2Fair leaves the generator’s internals and temporal process untouched. Our theoretical analysis establishes a finite-sample bound on the expected concept-level fairness discrepancy, accounting for finite observations and imperfect directional correction. Experiments demonstrate consistent fairness improvements across demographic attributes, prompt settings, and target distributions, with comparable prompt–video alignment and video quality, including on commercial black-box T2V models.
LLM Usage Statement
LLMs were used as assistive tools in the experiments and manuscript preparation. In the experimental pipeline, Qwen3-0.6B assisted with mapping prompts to occupation concepts and rewriting prompts to incorporate the selected protected-attribute value using a fixed instruction template, as described in § 5. OpenAI ChatGPT assisted with manuscript editing, language refinement and formatting. The authors retain responsibility for the research design, experimental execution, analysis, scientific claims, references, and final manuscript.
Additional Method Details and Analysis
This appendix supplements § 4 with formal history definitions, the proof of the conditional fairness guarantee, and the analysis of attribute selection.
History and Cold Start
Throughout this appendix, all counts and sequences are restricted to attribute-unspecified requests with valid detected outcomes (as defined in § 4.1); requests that are user-specified or yield an invalid detection are excluded from every index, count, and expectation below. For a global request index , let . Using the eligible index set from § 4.1, counts only attribute-unspecified requests with valid detected outcomes. User-specified attribute values are passed through unchanged and do not update Fair Memory. For requests with an existing state,
The projection used in Equation (6) is
It is unique because is closed, convex, and nonempty.
For cold-start requests, the controller samples . Because all attribute values are admissible, the distribution of selected attributes coincides with the target; the detected outcome is then a draw from the realized-outcome mixture , which equals only when the generator realizes every attribute exactly. The main theorem therefore does not rely on this coincidence and instead uses the directional condition on detected outcomes.
Proof of Approximate Conditional Approachability
Proof of Theorem 1. Fix a concept . Let satisfy for every one-hot and ; is valid on the simplex. Abbreviate
For each , also write
The empirical update over the concept-specific subsequence is, for ,
Proof of Approximate Conditional Approachability
Because ,
Taking conditional expectation and applying Assumption 1 and the definition of gives
For , . Summing the recursion from to therefore yields
Therefore,
The first outcome comes from the target-guided cold start, and the base case requires no directional condition. Jensen’s inequality and give the distance bound in the main text.
How Attribute Selection Enters the Guarantee
Nominal and true effects. The nominal effect used by the selector is the one-hot encoding of a candidate attribute:
Let contain the interaction history before request . The true effect denotes the conditional expected detected outcome when the prompt is rewritten with :
This is an analytical quantity, not a separate predictor queried by the controller.
Theorem 1 uses a local index , whereas the controller in § 4 uses the global request index . At request for concept with an existing memory state, the projection and residual and are the local quantities and for that concept. The controller has only one decision: choose and rewrite the prompt with that attribute. In the notation of § 4.2, its nominal effect is , while its true effect after rewriting and generation is . The theorem does not assume that the discrete selector in Equation (7) is solved exactly. To relate its slack to this selector, suppose that for each request after cold start and candidate attribute ,
Here bounds prompt-compliance error and bounds the residual norm ( is valid because both the empirical distribution and its projection lie in ). Assume also that the attribute selector is -optimal, meaning that the selected attribute satisfies
No reachability condition on the action set is needed: because and is the Euclidean projection of onto a closed convex , the projection inequality gives , and writing yields
Hence at least one attribute always offers a nonpositive correction score, and the directional condition holds with
The two terms account for prompt-compliance error and approximate attribute selection, respectively. For the tie-breaking rule of § 4.2, every tied attribute attains the minimum score exactly, so and .
Derivation of the directional slack. The compliance bound holds conditionally on the pre-decision history. Inserting the nominal effect and applying Cauchy–Schwarz to the prompt-compliance error gives, for any ,
Choosing and using -optimality together with the projection inequality above,
Thus Assumption 1 holds with
The compliance bound is an assumption about how reliably the generator realizes the selected attribute; it can be assessed from detected outcomes, without generating extra candidate videos.
Derivation of the Normalized Parity Score
This section derives the denominator of Equation (11) and relates to the fairness-discrepancy normalization of Teo et al. [30].
Normalization principle. Teo et al. [30] measure fairness by a discrepancy between a reference attribute distribution and the distribution realized by the generator, and observe that the raw discrepancy is not comparable across attributes with different numbers of groups , because its maximum attainable value grows with . Their fix is to normalize by the discrepancy evaluated at the absolutely-biased extreme point, i.e., , which places the score in regardless of . Fairmagan [11] instantiates this principle with the total-variation distance and a uniform reference , for which the absolutely-biased extreme point is any one-hot vertex and . Our applies the same normalization principle with an arbitrary target in place of the uniform reference.
Maximum attainable deviation from a general target. For a target on the -simplex , define the total-variation deviation . We show
attained at the one-hot vertex with . Since is a convex function of (a sum of absolute values of affine functions of ) and is a polytope, its maximum over is attained at a vertex, i.e., at some one-hot . At any vertex ,
using . This is maximized by choosing , giving Equation (30). Substituting and into the normalization of Teo et al. [30] recovers Equation (11) exactly, with the sign convention flipped so that (rather than ) denotes the fairest outcome.
Recovery of FairImagen and semantics of the uniform case. When for all , Equation (30) gives and every one-hot vertex attains it, so the worst case is exactly ”all generations collapse onto a single group,” matching the original FairImagen parity score. For a non-uniform , the vertices are no longer equidistant from the target: collapsing onto the already-favored group () yields a strictly smaller deviation than collapsing onto the least-favored group. Because is defined by the worst case, always denotes collapse onto , and intermediate values of should be read as a fraction of this worst-case deviation from rather than as a fraction of ”distance from a single-group outcome” in general.
Experimental Configuration
This section supplements § 5 with prompt preparation details and detection coverage.
Prompt preparation. Methods requiring training or calibration use the 50 seen occupations, disjoint from the 25 held-out occupations. Complex prompts are constructed from O*NET 31.0 task descriptions before evaluation and fixed across methods, without runtime regeneration or resampling.
Valid detection coverage. For method and protected attribute , we define detection coverage using the counts in Equation (10):
Failed detections remain in the denominator. Table 4 reports coverage in the Main setting. If no video yields a valid detection (), both fairness metrics are undefined and reported as dashes.
| Method | Gender () | Race () | Age () |
| Vanilla | 399/400 | 396/400 | 396/400 |
| FairGen | 400/400 | 396/400 | 396/400 |
| FairImagen | 397/400 | 400/400 | 390/400 |
| LightFair | 400/400 | 396/400 | 396/400 |
| Mem2Fair | 397/400 | 396/400 | 396/400 |
Table 4. Detection coverage in the Main setting: validly detected videos / generated videos, for each method and attribute task.
References
- [1]Gerdus Benadè, Aleksandr M. Kazachkov, Ariel D. Procaccia, and Christos-Alexandros Psomas. How to make envy vanish over time. In ACM Conference on Economics and Computation (EC), pages 593–610, 2018. 3DOI
- [2]Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency (FAccT), pages 1493–1504, 2023. 2DOI
- [3]David Blackwell. An analog of the minimax theorem for vector payoffs. Pacific Journal of Mathematics, 6(1):1–8, 1956. 1, 3DOI
- [4]Zefan Cai, Haoyi Qiu, Haozhe Zhao, Ke Wan, Jiachen Li, Jiuxiang Gu, Wen Xiao, Nanyun Peng, and Junjie Hu. From preferences to prejudice: The role of alignment tuning in shaping social bias in video diffusion models, 2026. 2arxiv.org/abs/2510.17247
- [5]José Cembrano, José Correa, Sebastian M. Griesbach, and Víctor Verdugo. Online proportional apportionment. In ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 4846–4860, 2023. 6
- [6]Jaemin Cho, Abhay Zala, and Mohit Bansal. DALL-Eval: Probing the reasoning skills and social biases of text-to-image generation models. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3043–3054, 2023. 2arxiv.org/abs/2202.04053
- [7]Yujin Choi, Jinseong Park, Hoki Kim, Jaewook Lee, and Saerom Park. Fair sampling in diffusion models through switching mechanism. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 2
- [8]Evgeni Chzhen, Christophe Giraud, and Gilles Stoltz. A unified approach to fair online learning via blackwell approachability. In Advances in Neural Information Processing Systems, 2021. 1, 3arxiv.org/abs/2106.12242
- [9]Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. RetinaFace: Single-shot multi-level face localisation in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 5DOI
- [10]Felix Friedrich, Manuel Brack, Lukas Struppek, Dominik Hintersdorf, Patrick Schramowski, Sasha Luccioni, and Kristian Kersting. Fair diffusion: Instructing text-to-image generation models on fairness, 2023. 1, 2arxiv.org/abs/2302.10893
- [11]Z. Fu, R. Brown, S. Shao, K. Rawal, E. D. Delaney, and C. Russell. Fairimagen: Post-processing for bias mitigation in text-to-image models. In Advances in Neural Information Processing Systems (NeurIPS), 2025. 1, 2, 5, 12arxiv.org/abs/2510.21363
- [12]Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5111–5120, 2024. 2arxiv.org/abs/2308.14761
- [13]Bingjie Gao, Xinyu Gao, Xiaoxue Wu, Yujie Zhou, Yu Qiao, Li Niu, Xinyuan Chen, and Yaohui Wang. The devil is in the prompts: Retrieval-augmented prompt optimization for text-to-video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3173–3183, 2025. 2arxiv.org/abs/2504.11739
- [14]Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In The Twelfth International Conference on Learning Representations, 2024. 1arxiv.org/abs/2307.10373
- [15]B. Han, Q. Xu, S. Bao, Y. Yang, K. Zi, and Q. Huang. Lightfair: Towards an efficient alternative for fair text-to-image diffusion via debiasing pre-trained text encoders. In Advances in Neural Information Processing Systems (NeurIPS), 2025. 1, 2, 5
- [16]Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. In Advances in Neural Information Processing Systems, pages 66923–66939. Curran Associates, Inc., 2023. 2arxiv.org/abs/2212.09611
- [17]Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chai, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21807–21818, 2024. 5arxiv.org/abs/2311.17982
- [18]Y. Jiang, W.-H. Li, Y. Zhang, M. Cai, and X. Yue. Fairgen: Enhancing fairness in text-to-image diffusion models via self-discovering latent directions. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 18411–18420, 2025. 1, 2, 5arxiv.org/abs/2412.18810
- [19]J. Li, L. Hu, J. Zhang, T. Zheng, H. Zhang, and D. Wang. Fair text-to-image diffusion via fair mapping. Proceedings of the AAAI Conference on Artificial Intelligence, 39(25):26256–26264, 2025. 1, 2arxiv.org/abs/2311.17695
- [20]Alexandra Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. Stable bias: Analyzing societal representations in diffusion models. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 2, 5arxiv.org/abs/2303.11408
- [21]Mohammad Nadeem, Shahab Saqib Sohail, Erik Cambria, Björn W. Schuller, and Amir Hussain. Gender bias in text-to-video generation models: A case study of Sora, 2025. 2arxiv.org/abs/2501.01987
- [22]Rishubh Parihar, Abhinya Bhat, Abhipsa Basu, Saswat Mallick, Jogendra Nath Kundu, and R. Venkatesh Babu. Balancing act: Distribution-guided debiasing in diffusion mod- els. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6668–6678, 2024. 2
- [23]Yujin Potter, Zhun Wang, Nicholas Crispino, Kyle Montgomery, Alexander Xiong, Ethan Y. Chang, Francesco Pinto, Yuqi Chen, Rahul Gupta, Morteza Ziyadi, Christos Christodoulopoulos, Bo Li, Chenguang Wang, and Dawn Song. VMDT: Decoding the trustworthiness of video foundation models, 2025. 2arxiv.org/abs/2511.05682
- [24]L. Qu, Z. Wang, N. Zheng, W. Wang, L. Nie, and T.-S. Chua. TTOM: Test-time optimization and memorization for compositional video generation. In The Fourteenth International Conference on Learning Representations (ICLR), 2026. 2arxiv.org/abs/2510.07940
- [25]Team Seedance, De Chen, Liyang Chen, Xin Chen, Ying Chen, Zhuo Chen, Zhuowei Chen, Feng Cheng, Tianheng Cheng, Yufeng Cheng, Mojie Chi, Xuyan Chi, Jian Cong, Qinpeng Cui, Fei Ding, Qide Dong, Yujiao Du, Haojie Duanmu, Junliang Fan, Jiarui Fang, Jing Fang, Zetao Fang, Chengjian Feng, Yu Gao, Diandian Gu, Dong Guo, Hanzhong Guo, Qushihan Guo, Boyang Hao, Hongxiang Hao, Haoxun He, Jiaao He, Qian He, Tuyen Hoang, Heng Hu, Ruoqing Hu, Yuxiang Hu, Jiancheng Huang, Weilin Huang, Zhaoyang Huang, Zhongyi Huang, Jishuo Jin, Ming Jing, Ashley Kim, Shanshan Lao, Yichong Leng, Bingchuan Li, Gen Li, Haifeng Li, Huixia Li, Jiashi Li, Ming Li, Xiaojie Li, Xingxing Li, Yameng Li, Yiying Li, Yu Li, Yueyan Li, Chao Liang, Han Liang, Jianzhong Liang, Ying Liang, Wang Liao, J. H. Lien, Shanchuan Lin, Xi Lin, Feng Ling, Yue Ling, Fangfang Liu, Jiawei Liu, Jihao Liu, Jingtuo Liu, Shu Liu, Sichao Liu, Wei Liu, Xue Liu, Zuxi Liu, Ruijie Lu, Lecheng Lyu, Jingting Ma, Tianxiang Ma, Xiaonan Nie, Jingzhe Ning, Junjie Pan, Xitong Pan, Ronggui Peng, Xueqiong Qu, Yuxi Ren, Yuchen Shen, Guang Shi, Lei Shi, Yinglong Song, Fan Sun, Li Sun, Renfei Sun, Wenjing Tang, Boyang Tao, Zirui Tao, Dongliang Wang, Feng Wang, Hulin Wang, Ke Wang, Qingyi Wang, Rui Wang, Shuai Wang, Shulei Wang, Weichen Wang, Xuanda Wang, Yanhui Wang, Yue Wang, Yuping Wang, Yuxuan Wang, Zijie Wang, Ziyu Wang, Guoqiang Wei, Meng Wei, Di Wu, Guohong Wu, Hanjie Wu, Huachao Wu, Jian Wu, Jie Wu, Ruolan Wu, Shaojin Wu, Xiaohu Wu, Xinglong Wu, Yonghui Wu, Ruiqi Xia, Xin Xia, Xuefeng Xiao, Shuang Xu, Bangbang Yang, Jiaqì Yang, Runkai Yang, Tao Yang, Yihang Yang, Zhixian Yang, Ziyan Yang, Fulong Ye, Bingqian Yi, Xing Yin, Yongbin You, Linxiao Yuan, Weihong Zeng, Xuejiao Zeng, Yan Zeng, Siyu Zhai, Zhonghua Zhai, Bowen Zhang, Chenlin Zhang, Heng Zhang, Jun Zhang, Manlin Zhang, Peiyuan Zhang, Shuo Zhang, Xiaohe Zhang, Xiaoying Zhang, Xinyan Zhang, Xinyi Zhang, Yichi Zhang, Zixiang Zhang, Haiyu Zhao, Huating Zhao, Liming Zhao, Yian Zhao, Guangcong Zheng, Jianbin Zheng, Xiaozheng Zheng, Zerong Zheng, Kuan Zhu, and Feilong Zuo. Seedance 2.0: Advancing video generation for world complexity, 2026. 1arxiv.org/abs/2604.14148
- [26]Sefik Ilkin Serengil and Alper Ozpinar. Lightface: A hybrid deep face recognition framework. In Innovations in Intelligent Systems and Applications Conference (ASYU), 2020. 5DOI
- [27]Preethi Seshadri, Sameer Singh, and Yanai Elazar. The bias amplification paradox in text-to-image generation. In Findings of the Association for Computational Linguistics: NAACL 2024, 2024. 2arxiv.org/abs/2308.00755
- [28]Xudong Shen, Chao Du, Tianyu Pang, Min Lin, Yongkang Wong, and Mohan Kankanhalli. Finetuning text-to-image diffusion models for fairness. In The Twelfth International Conference on Learning Representations, 2024. 1, 2arxiv.org/abs/2311.07604
- [29]Yongxu Sun, Michael Saxon, Ian Yang, Anna-Maria Gueorguieva, and Aylin Caliskan. VEAT quantifies implicit associations in text-to-video generator Soar and reveals challenges in bias mitigation, 2026. 2arxiv.org/abs/2601.00996
- [30]Christopher T. H. Teo, Milad Abdollahzadeh, and Ngai-Man Cheung. On measuring fairness in generative models. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 2, 5, 12arxiv.org/abs/2310.19297
- [31]Christopher T. H. Teo, Milad Abdollahzadeh, Xinda Ma, and Ngai man Cheung. Fairqueue: Rethinking prompt learning for fair text-to-image generation, 2024. 1, 2arxiv.org/abs/2410.18615
- [32]Team Wan, Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, Jiayu Wang, Jingfeng Zhang, Jingren Zhou, Jinkai Wang, Jixuan Chen, Kai Zhu, Kang Zhao, Keyu Yan, Lianhua Huang, Mengyang Feng, Ningyi Zhang, Pandleng Li, Pingyu Wu, Ruihang Chu, Ruili Feng, Shiwei Zhang, Siyang Sun, Tao Fang, Tianxing Wang, Tianyi Gui, Tingyu Weng, Tong Shen, Wei Lin, Wei Wang, Wei Wang, Wenmeng Zhou, Wente Wang, Wenting Shen, Wenyuan Yu, Xianzhong Shi, Xiaoming Huang, Xin Xu, Yan Kou, Yangyu Lv, Yifei Li, Yijing Liu, Yiming Wang, Yingya Zhang, Yitong Huang, Yong Li, You Wu, Yu Liu, Yulin Pan, Yun Zheng, Yuntao Hong, Yupeng Shi, Yutong Feng, Zeyinzi Jiang, Zhen Han, Zhi-Fan Wu, and Ziyu Liu. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. 1, 5arxiv.org/abs/2503.20314
- [33]Cong Wei, Quande Liu, Zixuan Ye, Quilin Wang, Xintao Wang, Pengfei Wan, Kun Gai, and Wenhu Chen. Univideo: Unified understanding, generation, and editing for videos. In The Fourteenth International Conference on Learning Representations, 2026. 1arxiv.org/abs/2510.08377
- [34]Jiwen Yu, Jianhang Bai, Yiran Qin, Quande Liu, Xintao Wang, Pengfei Wan, Di Zhang, and Xihui Liu. Context as memory: Scene-consistent interactive long video generation with memory retrieval, 2025. 2arxiv.org/abs/2506.03141
- [35]Cheng Zhang, Xuanbai Chen, Siqi Chai, Chen Henry Wu, Dmitry Lagun, Thabo Beeler, and Fernando De la Torre. ITI-GEN: Inclusive text-to-image generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3969–3980, 2023. 2arxiv.org/abs/2309.05569
- [36]Haonan Zhong, Wei Song, Tingxu Han, Maurice Pagnucco, Jingling Xue, and Yang Song. FAIRT2V: Training-free de-biasing for text-to-video diffusion models. arXiv preprint arXiv:2601.20791, 2026. 1, 2arxiv.org/abs/2601.20791
- [37]Jinsong Zhou, Yihua Du, Xinli Xu, Luozhou Wang, Zijie Zhuang, Yehang Zhang, Shuaibo Li, Xiaojun Hu, Bolan Su, and Ying-cong Chen. VideoMemory: Toward consistent video generation via memory integration, 2026. 2arxiv.org/abs/2601.03655


