Age Classifier Guidance in TimeMachine

• ACG is a training-time auxiliary supervision mechanism that improves fine-grained age controllability by aligning noisy latent states with target age using an L2 loss.
• It operates within the TimeMachine framework by predicting age attributes directly from diffusion states, avoiding instability from pixel-level reconstruction.
• Empirical results show that ACG enhances age accuracy subtly without affecting identity preservation, maintaining a balance between precision and training cost.

Age Classifier Guidance (ACG) is a training-time auxiliary age supervision mechanism introduced in the diffusion-based facial age-editing framework TimeMachine. Its purpose is to improve fine-grained age controllability without requiring explicit denoised-image reconstruction, while operating within a system whose broader objective is identity-preserving facial age editing. In TimeMachine, ACG estimates age-relevant attributes directly from noisy latent diffusion states and aligns them with the intended clean target through an $L_2$ objective, making it a latent-space supervision module rather than a sampling-time guidance rule in the classical diffusion sense (Mi et al., 15 Aug 2025).

1. Concept and motivation

ACG was proposed to address a specific weakness of prior diffusion-based face editing methods: they typically rely on the standard denoising or noise-prediction objective, so age control is indirect and often imprecise. When earlier methods add attribute supervision, they commonly do so by reconstructing or denoising an estimate of the clean image and then evaluating the attribute on that reconstructed output. TimeMachine argues that this is a poor fit for diffusion training, because a one-step or partial denoising estimate of the clean image is often inaccurate, which makes attribute supervision noisy and unstable. In facial age editing, where differences between ages can be subtle and tightly entangled with identity, expression, and texture, that instability directly impairs precise control (Mi et al., 15 Aug 2025).

The motivation for ACG is explicitly twofold. First, TimeMachine targets fine-grained age controllability, meaning accurate steering toward a precise target age rather than a coarse “young” versus “old” distinction. Second, it targets identity preservation, meaning that the edited face should remain recognizably the same person. These goals are in tension because age-related and identity-related factors are entangled in common latent representations. ACG is introduced as an additional supervisory mechanism that encourages the denoising process to preserve an age-consistent latent representation tied to the intended clean sample. Because it works in latent space rather than on reconstructed images, it is described as both more stable and lighter weight than pixel-space attribute supervision (Mi et al., 15 Aug 2025).

2. Placement within the TimeMachine architecture

TimeMachine is built on Stable Diffusion 1.5, with a denoising UNet conditioned by text, identity embeddings, and age embeddings through a Condition Projection Module and a Decoupled Multi-Cross-Attention mechanism. The upstream age/identity disentangling structure is expressed as

$$Out_{total} = Attn({Q}, {K}_t, {V}_t) + \sum_{i} \lambda_i \cdot Attn({Q}, {K}_i, {V}_i),$$

where $Q$ denotes hidden features from the UNet, $(K_t,V_t)$ are keys and values from text caption tokens, and $(K_i,V_i)$ are keys and values from condition-specific branches such as identity, age, and age-description. The scalar $\lambda_i$ controls the influence of each branch. This equation is not itself the ACG mechanism, but it defines the denoiser that ACG supervises (Mi et al., 15 Aug 2025).

Architecturally, ACG sits inside the training-time diffusion/denoising loop. It is not described as replacing any part of the main denoiser; instead, it is an auxiliary lightweight module attached to the diffusion training process. Figure 1(c) in the paper characterizes it as a module that “aligns age attributes in the latent space, enabling accurate age control without relying on pixel-level supervision.” Its key inputs are the noisy latent $z_t$, the timestep $t$, and denoiser-related noise information, written in the paper as $\boldsymbol{\epsilon}_\theta(z_t,t)$ or $\boldsymbol{\epsilon}_\theta(z_t,t,c)$. Functionally, ACG predicts the age of the corresponding clean latent or image $$Out_{total} = Attn({Q}, {K}_t, {V}_t) + \sum_{i} \lambda_i \cdot Attn({Q}, {K}_i, {V}_i),$$0 from that noisy diffusion state, without reconstructing the image itself (Mi et al., 15 Aug 2025).

3. Objective formulation and operational meaning

TimeMachine inherits the standard latent diffusion objective

$$Out_{total} = Attn({Q}, {K}_t, {V}_t) + \sum_{i} \lambda_i \cdot Attn({Q}, {K}_i, {V}_i),$$1

where $$Out_{total} = Attn({Q}, {K}_t, {V}_t) + \sum_{i} \lambda_i \cdot Attn({Q}, {K}_i, {V}_i),$$2 is the noised latent at timestep $$Out_{total} = Attn({Q}, {K}_t, {V}_t) + \sum_{i} \lambda_i \cdot Attn({Q}, {K}_i, {V}_i),$$3, $$Out_{total} = Attn({Q}, {K}_t, {V}_t) + \sum_{i} \lambda_i \cdot Attn({Q}, {K}_i, {V}_i),$$4 is sampled Gaussian noise, $$Out_{total} = Attn({Q}, {K}_t, {V}_t) + \sum_{i} \lambda_i \cdot Attn({Q}, {K}_i, {V}_i),$$5 is the conditioning signal, and $$Out_{total} = Attn({Q}, {K}_t, {V}_t) + \sum_{i} \lambda_i \cdot Attn({Q}, {K}_i, {V}_i),$$6 is the denoising network. The ACG term is added as

$$Out_{total} = Attn({Q}, {K}_t, {V}_t) + \sum_{i} \lambda_i \cdot Attn({Q}, {K}_i, {V}_i),$$7

where $$Out_{total} = Attn({Q}, {K}_t, {V}_t) + \sum_{i} \lambda_i \cdot Attn({Q}, {K}_i, {V}_i),$$8 is a pre-trained age classifier or predictor applied to the clean reference representation, and $$Out_{total} = Attn({Q}, {K}_t, {V}_t) + \sum_{i} \lambda_i \cdot Attn({Q}, {K}_i, {V}_i),$$9 is the lightweight latent-space age prediction module. The total objective becomes

$Q$0

with $Q$1 controlling the strength of the age constraint. The paper notes a capitalization inconsistency between $Q$2 and $Q$3, but treats them as the same term (Mi et al., 15 Aug 2025).

Several details are important for interpretation. The prose says that ACG uses the noisy latent $Q$4, timestep $Q$5, and the predicted noise $Q$6 to estimate the attribute of the corresponding clean image $Q$7, whereas the expectation in the equation includes $Q$8. The paper therefore contains a slight notation mismatch, but the intended meaning is clear: ACG infers age from denoiser state information in latent space. Although its name contains “classifier,” the equations and dataset formulation treat age numerically and optimize it with an $Q$9 loss. Operationally, ACG therefore behaves more like a regression-style auxiliary guidance module than a discrete class predictor (Mi et al., 15 Aug 2025).

These equations also clarify what ACG is not. It does not inject a classifier gradient into the reverse diffusion process in the style of classical classifier guidance, and the paper provides no modified reverse-step equation for inference. Instead, it changes the training objective of the denoiser by adding latent-space age supervision. In TimeMachine, “guidance” refers to shaping the denoiser’s learned representation so that noisy latent states remain age-predictive and age-consistent with the clean target (Mi et al., 15 Aug 2025).

4. Training regime, supervision sources, and dependence on HFFA

ACG is introduced only after an initial optimization stage. The experimental protocol is explicitly two-stage: Stage I performs initial training on the custom dataset as baseline training, and Stage II continues optimization by incorporating ACG. The paper does not describe ACG as a separate standalone pretraining phase; the most faithful reading is subsequent joint fine-tuning of the overall diffusion model with the additional ACG loss attached. Both stages use 8 NVIDIA H20 GPUs with batch size 8 per GPU. The paper repeatedly states that ACG improves age editing accuracy while only modestly increasing training cost, but does not report wall-clock overhead or FLOPs (Mi et al., 15 Aug 2025).

The relevant dataset is HFFA (High-quality Fine-grained Facial-Age dataset), constructed because of the lack of large-scale, high-quality facial age datasets. HFFA contains one million high-resolution images labeled with identity and facial attributes. In constructing it, the authors use MiVOLO to obtain two distinct forms of age supervision: a predicted numerical age from the final output layer and a latent age embedding $(K_t,V_t)$0 from the penultimate layer. The numerical age is used as the explicit age label, while the embedding is used for age-conditioning tokens and codebook construction. This is consequential for ACG because the $(K_t,V_t)$1 loss between age predictions implies continuous age supervision rather than one-hot age classes (Mi et al., 15 Aug 2025).

The paper also introduces an age codebook built by averaging age embeddings within each age cohort, with the stated goal of suppressing identity-specific variance and producing “purified” age-specific features. That purification matters indirectly for ACG: if age embeddings or labels retained too much identity information, the latent age-alignment objective would reinforce a contaminated target. HFFA is therefore important not only because of scale, but because it provides dense and relatively clean age annotation for latent-space age supervision (Mi et al., 15 Aug 2025).

5. Empirical effect and practical role

The paper provides one direct quantitative ablation isolating ACG’s contribution. In the condition-control ablation, the full model and the model without ACG perform as follows (Mi et al., 15 Aug 2025):

Setting	Age MAE	Face Similarity
Full model	3.840	0.67
Without ACG	3.915	0.67

This ablation shows that removing ACG slightly worsens age accuracy while leaving identity preservation unchanged in that experiment. The paper explicitly interprets this as meaning that ACG “does not drastically change the overall performance,” but “plays a subtle but important role in guiding the model toward more accurate and realistic age transformations during training.” The quantitative pattern supports that interpretation: ACG is not presented as the main driver of TimeMachine’s gains, but as a measurable refinement in age precision (Mi et al., 15 Aug 2025).

Its qualitative role is consistent with that result. The text suggests that ACG primarily improves realism and accuracy of subtle age-dependent cues rather than visibly altering identity structure. This matches the unchanged face-similarity score in the ablation. TimeMachine’s broader structural disentanglement already comes from the Condition Projection Module and the Decoupled Multi-Cross-Attention streams; ACG complements those mechanisms by adding a latent age-alignment signal during training. In practical terms, the module can be understood as asking whether the current noisy latent state and denoiser prediction already preserve the correct age of the clean target, and penalizing failures of that latent age consistency (Mi et al., 15 Aug 2025).

6. Conceptual interpretation, related methods, and open questions

A recurring misconception is to equate ACG with canonical diffusion classifier guidance. TimeMachine explicitly distinguishes the two. In traditional classifier guidance, a separately trained classifier $(K_t,V_t)$2 supplies gradients during sampling to steer reverse diffusion toward class $(K_t,V_t)$3. TimeMachine’s ACG is not used that way. It is a training-time latent supervision term, and despite its name it functions more as an auxiliary latent attribute supervision mechanism than as a classifier-gradient sampler modification. The paper also leaves several implementation details unspecified: the exact architecture of the “lightweight module” is not given; no layer counts, hidden sizes, or parameter counts are reported; the identity of $(K_t,V_t)$4 is not fully resolved in the text; and the numerical value or schedule of $(K_t,V_t)$5 is omitted. The modest ablation gain also leaves open the question of which age intervals or editing regimes benefit most strongly from ACG (Mi et al., 15 Aug 2025).

In broader methodological context, ACG aligns with a line of age-estimation research in which auxiliary signals are integrated structurally rather than used as brittle preprocessors. “A Unified Gender-Aware Age Estimation” argues that auxiliary classification should be unified with age estimation while preserving semantic distinction between tasks, rather than being handled by a sequential pipeline or naive joint target concatenation (Tian et al., 2016). “A Demographic Attribute Guided Approach to Age Estimation” uses auxiliary attribute branches and late fusion to improve age estimation, reinforcing the idea that side information can reshape age representations without replacing the main predictor (Cao et al., 2022). In a different safety-oriented setting, “Underage Detection through a Multi-Task and MultiAge Approach for Screening Minors in Unconstrained Imagery” shows that explicit threshold heads can outperform a pure age regressor for legally consequential boundaries, which suggests that decision-aligned age supervision often benefits from auxiliary heads rather than a single undifferentiated age output (Gaul et al., 12 Jun 2025).

Taken together, these adjacent results suggest a consistent interpretation of ACG: it is not a standalone age-editing engine, nor a generic classifier-guidance algorithm, but a structured auxiliary supervision mechanism for stabilizing and sharpening age information inside a larger age-conditioned model. Within TimeMachine specifically, its role is narrow but technically precise: to maintain an age-faithful latent denoising trajectory while leaving identity structure to the framework’s explicit disentangling architecture (Mi et al., 15 Aug 2025).