# Controllable Image Generation with Composed Parallel Token Prediction

Source: [Controllable Image Generation with Composed Parallel Token Prediction](<https://www.sota2.com/research/paper/69d5e1067d0172140ec75a5d>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-04-09T05:00:04.534Z

Authors: Jamie Stirling, Noura Al-Moubayed, Chris G. Willcocks, Hubert P. H. Shum

Published: 2026-04-07T11:33:52.000Z

## Abstract

Conditional discrete generative models struggle to faithfully compose multiple input conditions. To address this, we derive a theoretically-grounded formulation for composing discrete probabilistic generative processes, with masked generation (absorbing diffusion) as a special case. Our formulation enables precise specification of novel combinations and numbers of input conditions that lie outside the training data, with concept weighting enabling emphasis or negation of individual conditions. In synergy with the richly compositional learned vocabulary of VQ-VAE and VQ-GAN, our method attains a $63.4\\%$ relative reduction in error rate compared to the previous state-of-the-art, averaged across 3 datasets (positional CLEVR, relational CLEVR and FFHQ), simultaneously obtaining an average absolute FID improvement of $-9.58$. Meanwhile, our method offers a $2.3\\times$ to $12\\times$ real-time speed-up over comparable methods, and is readily applied to an open pre-trained discrete text-to-image model for fine-grained control of text-to-image generation.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Compositional Image Generation on FFHQ 1 Component](<https://www.sota2.com/research/sota/compositional-image-generation-on-ffhq-1-component.md>): Compositional Image Generation; FFHQ 1 Component; Error Rate (%): 0.22.

- [Compositional Image Generation on Relational CLEVR 1 Component](<https://www.sota2.com/research/sota/compositional-image-generation-on-relational-clevr-1-component.md>): Compositional Image Generation; Relational CLEVR 1 Component; Error Rate (%): 21.84.

- [Compositional Image Generation on Positional CLEVR 1 Component](<https://www.sota2.com/research/sota/compositional-image-generation-on-positional-clevr-1-component.md>): Compositional Image Generation; Positional CLEVR 1 Component; Error (%): 0.7.

- [Compositional Image Generation on FFHQ 2 Components](<https://www.sota2.com/research/sota/compositional-image-generation-on-ffhq-2-components.md>): Compositional Image Generation; FFHQ 2 Components; Error Rate (%): 0.62.

- [Compositional Image Generation on FFHQ (3 Components)](<https://www.sota2.com/research/sota/compositional-image-generation-on-ffhq-3-components.md>): Compositional Image Generation; FFHQ (3 Components); Error (%): 0.82.

- [Compositional Image Generation on Relational CLEVR 2 Components](<https://www.sota2.com/research/sota/compositional-image-generation-on-relational-clevr-2-components.md>): Compositional Image Generation; Relational CLEVR 2 Components; Error Rate: 56.94.

- [Compositional Image Generation on Relational CLEVR 3 Components](<https://www.sota2.com/research/sota/compositional-image-generation-on-relational-clevr-3-components.md>): Compositional Image Generation; Relational CLEVR 3 Components; Error Rate (%): 85.7.

- [Compositional Image Generation on Positional CLEVR 2 Components](<https://www.sota2.com/research/sota/compositional-image-generation-on-positional-clevr-2-components.md>): Compositional Image Generation; Positional CLEVR 2 Components; Error Rate (%): 1.82.

- [Compositional Image Generation on Positional CLEVR 3 Components](<https://www.sota2.com/research/sota/compositional-image-generation-on-positional-clevr-3-components.md>): Compositional Image Generation; Positional CLEVR 3 Components; Error Rate (%): 4.96.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2604.05730>)

[PDF](<https://arxiv.org/pdf/2604.05730.pdf>)
