Constrained Image-Synthesis Lab

Projects

Diffusion Models for Flexible Packaging

Our current research initiative exploring generative models for industrial design applications.

Under this umbrella, we are working on three sub-projects addressing physical and perceptual constraints in print-ready image synthesis.

RobustSep

PPP-conditioned RGB to CMYK separation with constraint refinement

Problem

Separation is conditioned on a Process-and-Press Prior (PPP) and refined through explicit constraints rather than treated as unconstrained translation.

Constraints Enforced

Conservative ink-limit logic (TAC and per-channel caps), overprint sanity, neutral/dark stability, separation regularity.

Method

Patch-aware (16×16), proposes multiple candidates per patch, applies robustness checks under parametric perturbations (tone curves, ink strength). Outputs CMYK result and concise decision report (risk flags and constraint tightness).

Evaluation

TAC violation rate, per-channel cap adherence, overprint stability, neutral preservation (ΔE), robustness score under PPP perturbations.

Citation

Retained for private usage. Will be published after redaction.

Deep Reinforcement Learning Halftoning

Halftoning under print constraints via deep reinforcement learning

Problem

Halftoning framed as constrained optimization with deep reinforcement learning as comparator to classical baselines.

Constraints Enforced

Per-channel frequency bounds, dot gain limits, moiré avoidance, local tone linearity.

Baselines

JJN error diffusion, void-and-cluster blue-noise ordered dithering, Direct Binary Search.

Evaluation

Constraint pass rate, robustness under gain variation, perceptual similarity (SSIM, LPIPS), per-channel frequency analysis.

Citation

Retained for private usage. Will be published after redaction.

Inference-Time Loss-Guided Colour Preservation in Diffusion Sampling

Inference-time constraint enforcement for perceptual colour fidelity

Problem

Maintaining perceptual colour fidelity when generative processes introduce stochastic variations that can drift from intended colour specifications.

Constraints Enforced

Perceptual colour drift bounds (ΔE metrics), region-specific preservation, semantic colour consistency across generated outputs.

Method

Inference-time losses and regularizers with explicit monitoring of colour drift and perceptual reconstruction error.

Evaluation

Perceptual colour drift (ΔE), region preservation metrics, semantic consistency scores, constraint violation frequency.

Cite this work Pre-print · Out now
@article{colourperception2025,
  title={Inference-Time Loss-Guided Colour Preservation in Diffusion Sampling},
  author={Ahuja, Angad Singh and Patel, Poorva and Anandh, Aarush Ram},
  journal={arXiv preprint},
  year={2025},
  note={Pre-print, out now}
}