CompSlider: Optimizing Generative & Convex Models

• CompSlider is a framework of methods that enable fine-grained control over multiple attributes in generative models and efficient composite convex optimization.
• The generative model variant uses a diffusion-transformer to jointly embed slider tokens, achieving lower entanglement (14%) and high consistency (90.9%).
• The optimization approach alternates expensive gradient evaluations with efficient subgradient steps, offering optimal oracle complexities for composite convex problems.

CompSlider refers to several influential techniques across optimization, engineering, and generative models, where “sliding” or “slider” indicates either physical variable motion, algorithmic decoupling of computations, or fine-grained user manipulation of attributes. Notable instantiations appear in composite convex optimization, image and video generative modeling, and sliding-cable analysis. This article surveys major CompSlider methodologies, their theoretical foundations, architectures, losses, and evaluation paradigms, with detailed references to canonical works.

1. CompSlider for Multi-Attribute Control in Generative Models

The most recent development under the CompSlider name is “CompSlider: Compositional Slider for Disentangled Multiple-Attribute Image Generation” (Zhu et al., 31 Aug 2025). CompSlider here is a conditional-prior generator designed to enable reliable, simultaneous, and disentangled control over multiple semantic attributes (e.g., age, smile) in text-to-image generation, with generalization to video. Unlike methods such as ConceptSlider (Gandikota et al., 2023), which train separate adapters for each attribute—leading to attribute entanglement during composition—CompSlider jointly embeds all slider values as tokens and leverages a diffusion-transformer (DiT) to predict the conditional prior to feed a frozen backbone diffusion model.

Architecture Overview

• Foundation: A “frozen” foundation T2I model (e.g., a CLIP-conditioned diffusion backbone) remains untouched during CompSlider training and inference.
• Inputs: User provides
  • Text prompt $x^{\mathcal T}$, encoded to $c^{\mathcal T}$
  • $N$-dimensional slider vector $v^{\mathcal S}\in[0,1]^N$, embedded as $N$ tokens $c^{\mathcal S}$
• DiT Conditioning: A lightweight DiT consumes noise-corrupted condition $c^{\mathcal I}_t$, slider tokens $c^{\mathcal S}$, and text tokens $c^{\mathcal T}$, and predicts $c^{\mathcal I}_0$.
• Generation: The predicted $c^{\mathcal T}$0 and $c^{\mathcal T}$1 are passed to the frozen U-Net, generating the image.

Objective Functions

Let DiT$c^{\mathcal T}$2.

• Diffusion loss:

$c^{\mathcal T}$3

• Disentanglement loss:

Pairs of slider settings $c^{\mathcal T}$4 are used; an auxiliary MLP predicts quantized bucketed differences, penalized by cross-entropy between predicted bucket and true difference.

• Structure loss:

For small per-attribute changes $c^{\mathcal T}$5, enforce $c^{\mathcal T}$6 proximity between predicted conditions under $c^{\mathcal T}$7 and $c^{\mathcal T}$8.

• Total loss:

$c^{\mathcal T}$9

Attribute Composability

Unlike vector combination or explicit addition, token-wise cross-attention in the DiT fuses all $N$0 attributes, learned jointly for disentanglement. The system supports arbitrary-real-valued manipulation of each attribute in a single inference call.

Performance and Metrics

Key evaluation metrics proposed:

• Continuity: Fraction of monotonic attribute scores (DeepFace predictivity)
• Scope: Score change from minimum to maximum slider value
• Consistency: Identity preservation rate (DeepFace recognition)
• Entanglement: Fractional change in unintended attributes during single-attribute manipulation

CompSlider achieves a lower entanglement (14%) and higher consistency (90.9%) compared to ConceptSlider, with improvements in continuity and scope as well (Zhu et al., 31 Aug 2025). Human preference experiments and ablations confirm the effectiveness of incorporating both disentanglement and structure losses.

2. CompSlider in Composite Convex Optimization

CompSlider, or the Composite Sliding Method, appears as a class of algorithms for composite convex optimization. In this context, CompSlider alternates between computationally costly and cheap oracles to efficiently attain optimal rates for minimization of the form $N$1, where $N$2 is smooth (and possibly strongly convex) and $N$3 is non-smooth or composite (Lan, 2014, Beznosikov et al., 2019, Dvinskikh et al., 2019).

CompSlider (Gradient Sliding) (Lan, 2014)

CompSlider achieves optimal oracle complexities by performing expensive $N$4 steps infrequently, “sliding” many inner subgradient steps for $N$5 and $N$6 terms, allowing total $N$7 calls to be $N$8 for general convex objectives and $N$9 for strongly convex $v^{\mathcal S}\in[0,1]^N$0, while total g/h subgradient calls scale as $v^{\mathcal S}\in[0,1]^N$1 and $v^{\mathcal S}\in[0,1]^N$2, respectively.

Algorithmic Structure

• Outer loop: Occasional affine expansion using $v^{\mathcal S}\in[0,1]^N$3
• Inner (prox-sliding) loop: Many subgradient/prox/zeroth-order (for nonsmooth terms)
• Prox-sliding subroutine structured with martingale weights $v^{\mathcal S}\in[0,1]^N$4 for acceleration and variance control.

This permits stochastic extensions, high-probability guarantees, and saddle-point generalizations (Lan, 2014).

Derivative-Free Composite Sliding (zoSA) (Beznosikov et al., 2019)

The method combines a stochastic zeroth-order oracle (ZOO) for the nonsmooth term $v^{\mathcal S}\in[0,1]^N$5 and a first-order oracle for the smooth $v^{\mathcal S}\in[0,1]^N$6. Convergence rates match first-order methods up to $v^{\mathcal S}\in[0,1]^N$7 factors. Key elements include:

• Spherical smoothing of the nonsmooth $v^{\mathcal S}\in[0,1]^N$8 via randomized finite differences
• Sliding/proximal steps leveraging both oracles
• Decentralized optimization application over graphs, achieving optimal communication complexity bounds per node.

Strongly Convex and Variance-Reduced Extensions (Dvinskikh et al., 2019)

CompSlider, with Catalyst acceleration and variance-reduced inner solvers (SVRG, SAGA, Katyusha), achieves:

$v^{\mathcal S}\in[0,1]^N$9

[ O\left(\sqrt{\frac{mL_g}{\mu}}\log\frac1