Keeping Cultural Detail Intact When an Image Generator Is Under Load

When a heritage brand launches a product or an influencer campaign goes viral, thousands of people may ask an image generator at once to turn a historical garment or embroidery pattern into a modern, wearable design. The usual engineering response to a traffic spike is to cut computation per request: fewer denoising steps, a smaller model, lower resolution. For generic images this costs little. For cultural designs it may cost almost everything, because what separates faithful heritage from a generic "Chinese-style" look is concentrated in small, rare details. Examples include the cut of a collar, the overlap of a front panel, and the shape of a sleeve that marks a particular dynasty. Our hypothesis is that these fine-grained, period-defining attributes are the first thing lost when computation is cut, well before overall image quality looks worse. If that is true, a load-shedding policy should decide where to cut by measuring which cultural attributes each request's output is likely to lose, not by generic image quality.

We propose to measure this "fidelity cliff" directly, and then to build a serving controller that protects it. The controller watches queue delay and sets an overall compute budget. It then spends that budget request by request, choosing model tier, step count, and whether to run a targeted refinement pass. Each choice is made to keep a floor on the attributes that experts mark as period-defining. The claim is testable in both directions. If protected attributes degrade no faster than generic quality under compute cuts, uniform degradation is enough and the controller is unnecessary. We treat that outcome as the first decisive check, not as something to argue around.

Where the idea comes from

The systems side supplies the mechanism. Adjusting Matching Algorithm to Adapt to Dynamic Subscriptions in Content-Based Publish/Subscribe Systems shows that when hardware cannot be added fast enough to absorb micro-bursts, a system can adapt its algorithm instead. Their broker skips rarely used subscription attributes under load, and a controller based on proportional-integral-derivative (PID) feedback sets the skip rate from observed latency. It halves average latency, from 57.11 ms to 27.32 ms, with under 0.1% false positives and no false negatives. We reuse the feedback-controller design. We reverse its core assumption, though. That broker sheds precisely the long-tail work. In cultural transcreation, meaning adapting a historical design into a modern functional product, the long tail is where the value lies. Our analogue of its "zero false negatives" guarantee is a floor on protected attributes.

The second systems source is Online Job Scheduling in Distributed Machine Learning Clusters. It admits and resizes machine-learning training jobs with resource prices that rise exponentially as the cluster fills. It treats job value as a fixed abstract function. We borrow the pricing idea to turn the global budget into a per-request price. We replace abstract value with measured, attribute-level retention. Learning Agent Routing From Early Experience routes queries between cheap and expensive model pipelines on generic tasks. It is the closest method competitor, and we run it as a baseline rather than claim it cannot work.

The heritage side supplies the motivation and the evaluation. Hanfu-Bench: A Multimodal Benchmark on Cross-Temporal Cultural Understanding and Transcreation provides 496 expert-annotated garment sets (1,192 images) labeled by period and by component: sleeve, front panel (jin), collar, bottoms, outerwear. It reports that the best model reaches only a 42% human-judged transcreation success rate. That means there is little quality margin to spend under load. Design studies of Huangmei embroidery, Visual Transformation of Textile Intangible Cultural Heritage Patterns in the Design of Cultural and Creative Products, and of opera motifs, Research on the Development and Visual Image of Opera Culture Creative Products, argue qualitatively that products fail when they copy surface shape and lose defining detail. These studies motivate which attributes to protect. They supply no measurements. Innovative Integration of Digital Media Art Design and Traditional Chinese Cultural Elements -- A Case Study of the "Florasis·Miao Impressions" Series documents interactive heritage media deployed at consumer scale. It says nothing about the serving infrastructure behind it, which is the gap we target. The research file found no prior work joining these two directions.

Core Claim

Under bursty load, cutting computation per request in cultural image transcreation degrades expert-defined, period-defining garment attributes much faster than it degrades generic quality scores. A controller that allocates computation by predicted per-attribute retention can hold protected-attribute fidelity near its unloaded level at the same latency target and hardware cost where uniform degradation, generic routing, and caching cannot.

Research Program

The work has three parts.

The first part is a sensitivity map. We take an open-weight image generation pipeline that exposes several knobs: two or three model sizes, a range of denoising step counts, output resolution, and an optional masked refinement pass over one garment region. We run it over Hanfu-Bench transcreation prompts across the full grid of settings. For every output we score each annotated component for retention. We also compute a generic score, such as CLIP-based prompt alignment and an aesthetic predictor.

Scoring needs a reliable attribute checker. Hanfu-Bench itself reports that closed vision-language models fall about 10% short of experts on cultural understanding, so an off-the-shelf judge cannot be trusted. We will therefore fine-tune a component-level classifier on Hanfu-Bench annotations. We will validate it against a held-out set of human expert judgments, report agreement, and use it only where agreement is acceptable. This validation is a gate: if the checker is unreliable on some component, that component is reported only through human evaluation.

The second part is the controller. An outer PID loop, as in the broker work, reads queue delay and sets a compute budget $B_t$ for the next interval. An inner allocator gives each queued request $i$ a knob setting $k_i$ from a small menu. It maximizes expected retention of that request's protected attributes subject to the budget:

$$\max_{\{k_i\}} \sum_i \min_{a \in P_i} \hat r_{a}(k_i, x_i) \quad \text{s.t.} \quad \sum_i c(k_i) \le B_t .$$

Here $P_i$ is the set of period-defining attributes for the requested garment, taken from the expert annotations. The term $\hat r_a(k, x)$ is the predicted probability that attribute $a$ survives under setting $k$ for input $x$. A lightweight predictor trained on the sensitivity map supplies it. The term $c(k)$ is the GPU cost of setting $k$. The inner minimum encodes a floor rather than an average: a request is only as good as its most fragile protected attribute. We solve the problem greedily, using a single price per unit of compute in the spirit of the exponential-pricing scheduler. The refinement pass matters here. It lets the controller spend a small amount of extra compute on the one region, often the collar or front panel, that a cheap setting would otherwise lose. This is cheaper than upgrading the whole request.

The third part is a load study. No traffic traces for heritage campaigns exist. We will therefore drive the system with arrival processes whose burst statistics come from traces the research file documents. From the Huawei serverless request trace used in Green or Fast? Learning to Balance Cold Starts and Idle Carbon in Serverless Computing we take inter-arrival structure. From the stock-exchange trace we take burst magnitudes: minute-to-minute swings over 60% and spikes up to 767%. We rescale both to the throughput of a fixed GPU pool. This is a stand-in for real campaign traffic, and we say so. Results are reported per trace so that conclusions do not rest on one burst shape.

Baselines

The naive reference points come first. One is a statically over-provisioned pool that always runs the highest setting. It defines the fidelity ceiling and the cost ceiling, and it tells us how much quality a policy gives up at a given budget. The other is a threshold autoscaler that adds GPU replicas when utilization passes 80%, charged a realistic cold-start delay. Elastic Resource Allocation for Volatile Workloads describes this kind of trigger as prone to oscillation and lag. It shows what adding capacity alone achieves on these bursts.

The strongest simple competitor is uniform degradation under the same PID loop. It lowers step count or model tier equally for every request as the budget shrinks. It shares our outer controller exactly, so it isolates whether per-request, attribute-aware allocation adds anything.

The strongest existing method is a learned router in the style of Learning Agent Routing From Early Experience. It picks among the same knob settings using a predicted generic quality score, retrained on our sensitivity map with the same data and tuning budget as our predictor.

The key ablation is our allocator with $\hat r_a$ replaced by a generic quality predictor. Only the objective differs in this ablation. It tests whether cultural attribute signals, as opposed to per-request adaptivity in general, drive any gain. An oracle allocator that knows true retention bounds the achievable result.

All policies run on the same GPU pool, the same prompts, the same arrival traces, and the same latency objective. Each gets an equal hyperparameter search.

There is no reported result on Lacuna for attribute-fidelity-aware transcreation serving under load. For the underlying generation task, the reference number is Hanfu-Bench's best-model transcreation success rate of 42%, from human evaluation. For elastic control, the analogous systems number is Ada-Rein's reduction in average event latency from 57.11 ms to 27.32 ms, from the broker paper. That comes from a different domain and is not a target for us. Beating a baseline means one of two things, with human-evaluated success on a sampled subset confirming the attribute-checker numbers. Either we achieve higher protected-attribute retention at matched p95 latency and matched GPU-hours, or we achieve matched retention at clearly lower cost or fewer latency-target violations.

Why not just…?

Why not just add GPUs when load rises, or over-provision? Over-provisioning works, but it pays full cost during quiet periods. The serverless study measured that a static keep-alive policy wastes energy on idle capacity: a learned policy cut idle keep-alive carbon by 77.08% and cold starts by 51.69%. Autoscaling responds in seconds to minutes, and the burst traces we replay change faster than that. Both are run as baselines. The static pool gives the cost of full fidelity, and the autoscaler shows whether capacity alone meets the latency target during spikes.

Why not just degrade every request uniformly, which is simpler? This is the real threat. If protected attributes degrade at the same rate as generic quality, uniform degradation under the same controller is enough, and our allocator is extra complexity. The sensitivity map settles this before any controller is built. We compare, per component, the knob setting at which retention drops by a fixed margin against the setting at which the generic score drops by the same margin. If there is no gap, we stop and report that.

Why not just pre-generate designs offline and serve the closest cached one? Caching is attractive because heritage motifs are fixed. The design literature argues that stamped, repeated patterns produce homogenized products that consumers reject. That argument is qualitative, though, so we test caching directly. We include a nearest-neighbor cache over pre-generated outputs as an extra arm, and we measure how often a cached result satisfies the request's specific modern-function constraints according to the attribute checker plus human raters. If caching matches us on common requests, we will fold it into the controller as its cheapest setting rather than compete with it.

Why not just use a generic learned router? It is a baseline, and the generic-predictor ablation is its fairest form. If it matches our allocator, then the cultural attribute signal adds nothing beyond adaptivity. That would be a clean negative result.

Decisive Checks

Three results decide the project. First, the sensitivity map must show that at least some period-defining components drop off at settings where generic scores are still flat. Without that, the premise fails. Second, the attribute checker must agree well enough with expert human judgment on held-out outputs to be used as a metric. Otherwise all fidelity claims rest on human evaluation alone, and the load study shrinks accordingly. Third, under replayed bursts at matched latency and GPU-hours, the fidelity-aware allocator must beat the generic-predictor ablation on worst-attribute retention and on human-judged success. The gain should come mostly during bursts, since everyone runs the full setting when load is low. If the gain appears only against uniform degradation and not against the ablation, the lesson is about per-request adaptivity, not cultural fidelity, and we will report it that way.