CompSlider: Compositional Slider for Disentangled Multiple-Attribute Image Generation¶

Conference: ICCV 2025 arXiv: 2509.01028 Code: None Area: Image Generation Keywords: attribute disentanglement, slider control, text-to-image generation, conditional prior, multi-attribute manipulation

TL;DR¶

This paper proposes CompSlider, a compositional slider model that generates conditional priors to enable simultaneous, independent, and fine-grained control over multiple attributes in T2I foundation models. It addresses inter-attribute entanglement via a disentanglement loss and a structural consistency loss.

Background & Motivation¶

In text-to-image (T2I) generation, controlling the intensity of image attributes (e.g., age, degree of smile) through text prompts alone is inherently imprecise. Slider-based generation methods (e.g., ConceptSliders, PromptSlider) have emerged to allow users to continuously adjust attributes via sliders. However, existing methods train a separate adapter for each attribute and overlook inter-attribute entanglement:

Attribute entanglement: Composing sliders in different orders yields different results — e.g., applying "smile" before "age" produces a different outcome than the reverse order.

Structural inconsistency: Adjusting one attribute inadvertently alters unrelated factors such as background or hairstyle.

Poor scalability: \(N\) attributes require \(N\) forward passes, imposing significant computational overhead.

Method¶

Overall Architecture¶

CompSlider replaces the role of the CLIP image encoder in the T2I foundation model. It takes user-defined slider values and a text prompt as input, and outputs an image condition \(\bm{c}^{\mathcal{I}}\) as a multi-attribute prior, which is fed into the base diffusion model for image generation:

\[\bm{c}^{\mathcal{I}} = \text{CompSlider}(\bm{c}^{\mathcal{S}}, \bm{c}^{\mathcal{T}})\]

where \(\bm{c}^{\mathcal{S}}\) denotes slider embeddings and \(\bm{c}^{\mathcal{T}}\) denotes T5 text tokens. The base model requires no fine-tuning.

Key Designs¶

DiT as the CompSlider backbone: A Diffusion Transformer (DiT) is adopted, using a reparameterization trick to directly predict the clean image condition \(\bm{c}_0^{\mathcal{I}}\) rather than noise. Since the image condition is a 1024-dimensional vector with no need for downsampling, DiT is more suitable than U-Net. The model consists of 10 DiT blocks, taking 128 text tokens and 16 slider tokens as input, with a total of 277M parameters.
Slider value embedding: Attribute scores are obtained via pretrained attribute classifiers and normalized to \([0,1]\). Positional encoding (sinusoidal encoding) maps continuous slider values to vectors \(\bm{p}^{\mathcal{S}} \in \mathbb{R}^{N \times \frac{dim}{2}}\), and learnable category embeddings \(\bm{w} \in \mathbb{R}^{N \times \frac{dim}{2}}\) are introduced to distinguish different attributes. The final slider embedding is their concatenation: \(\bm{c}^{\mathcal{S}} = [\bm{p}^{\mathcal{S}}, \bm{w}]\).
Random attribute combination training strategy: A key innovation is that the method does not rely on paired data (i.e., images of the same subject at varying attribute intensities). Instead, randomly sampled attribute value combinations \(\bm{v}^{\mathcal{S}*}\) are introduced during training, ensuring the model generalizes to arbitrary combinations rather than only the co-occurrence patterns observed in training data.

Loss & Training¶

The total loss comprises three terms: \(\mathcal{L} = \mathcal{L}_{\text{diff}} + \mathcal{L}_{\text{st}} + \mathcal{L}_{\text{clss}}\)

Diffusion loss \(\mathcal{L}_{\text{diff}}\): Ensures the generated condition aligns with the output distribution of the CLIP image encoder: \(\mathcal{L}_{\text{diff}} = \mathbb{E}[\|\bm{c}_0^{\mathcal{I}} - \text{DiT}(\bm{c}_t^{\mathcal{I}}, \bm{c}^{\mathcal{S}}, \bm{c}^{\mathcal{T}}, t)\|^2]\)
Disentanglement loss \(\mathcal{L}_{\text{clss}}\): An MLP classifier is trained to recover attribute differences from the conditional discrepancy between original and randomly combined attribute outputs. The difference is discretized into \(B=20\) bins and supervised with a cross-entropy loss.
Structural loss \(\mathcal{L}_{\text{st}}\): When the attribute difference satisfies \(|\Delta v_i| \leq 0.1\), the L2 distance between the two conditional outputs is constrained to preserve local structural consistency.

Training data comprises approximately 3 million images. The diffusion and structural losses train the DiT, while the disentanglement loss jointly trains both the DiT and the MLP classifier.

Key Experimental Results¶

Main Results¶

Quantitative comparison on human-related sliders (300 prompts × 5 attributes × 5 values = 7,500 images):

Method	Cont.%↑	Cons.%↑	Scope%↑	Entang.%↓	LPIPS↓	CLIP↑
Prompt2Prompt	-	88.47	49.46	28.99	0.19	4.15
PromptSlider	61.17	80.23	46.25	24.31	0.10	4.79
ConceptSlider	73.41	83.17	54.43	27.22	0.16	5.76
CompSlider	81.07	90.95	59.02	14.04	0.12	6.20

A/B user study on non-human sliders (Vector Style + Scene Complexity): CompSlider achieves a user preference of 54.66% vs. 34.16% for ConceptSlider.

Ablation Study¶

Ablation of the disentanglement loss and structural loss:

\(\mathcal{L}_{\text{diff}}\)	\(\mathcal{L}_{\text{clss}}\)	\(\mathcal{L}_{\text{st}}\)	Cont.%↑	Cons.%↑	Scope%↑	Entang.%↓
✓			68.96	63.21	42.06	36.68
✓	✓		76.49	49.29	63.27	19.87
✓	✓	✓	81.07	90.95	59.02	14.04

Key Findings¶

Using only the diffusion loss results in an entanglement rate as high as 36.68%; adding the disentanglement loss reduces it to 19.87%, but structural consistency collapses to 49.29%.
The structural loss substantially recovers consistency from 49.29% to 90.95% (+41.66%) and further reduces entanglement to 14.04%.
CompSlider supports controlling all sliders in a single forward pass, offering substantially better inference efficiency than per-attribute methods.

Highlights & Insights¶

Core innovation: Operating in the latent space of conditional priors eliminates the need to fine-tune the base model, significantly reducing training and inference costs.
No paired data required: Disentanglement is learned via randomly sampled attribute combinations during training, elegantly circumventing the difficulty of acquiring paired data of the same subject at varying attribute intensities.
Four new evaluation metrics: Continuity, Scope, Consistency, and Entanglement are proposed, providing a more comprehensive assessment of slider-based generation quality than LPIPS/CLIP alone.
The approach is extensible to video generation.

Limitations & Future Work¶

The method relies on pretrained attribute classifiers to obtain ground-truth slider values; classifier quality directly impacts training.
The attribute set is closed (16 predefined attributes), with no support for open-domain attributes.
Compatibility with more recent diffusion models (e.g., SDXL, SD3) has not been validated.
Automated evaluation metrics for non-human attributes are absent; assessment relies solely on user studies.

ConceptSliders and PromptSlider are direct predecessors, employing LoRA adapters and textual inversion for single-attribute sliders, respectively.
eDiff-I provides the foundational framework for conditional image priors.
The disentanglement strategy offers inspiration for other multi-condition controllable generation tasks, such as simultaneously controlling composition, style, and content.

Rating¶

Novelty: ⭐⭐⭐⭐ Compositional sliders operating in the conditional prior space is a novel perspective, and the disentanglement loss design is elegant.
Experimental Thoroughness: ⭐⭐⭐⭐ Evaluation covers both human and non-human attributes, new metrics are proposed, and ablation studies and extensions are included.
Writing Quality: ⭐⭐⭐⭐ Motivation is clearly articulated, the method is described in detail, and figures and tables are informative.
Value: ⭐⭐⭐⭐ Addresses a practical problem of simultaneous multi-attribute control with clear application scenarios.