Beyond Starry Night: Shortcut-Aware Control-State Planning for Artist-Grounded Text to Image Generation
Summary
This paper introduces Atelier, a method that plans explicit control states before generation to prevent text-to-image models from falling back on artist-name shortcuts, and presents ArtIntentBench for evaluating artist-grounded style control.
View Cached Full Text
Cached at: 08/12/26, 08:22 AM
Paper page - Beyond Starry Night: Shortcut-Aware Control-State Planning for Artist-Grounded Text to Image Generation
Source: https://huggingface.co/papers/2608.06751 Beyond Starry Night: Shortcut-Aware Control-State Planning for Artist-Grounded Text-to-Image Generation
📄 arXiv:https://arxiv.org/abs/2608.06751
TL;DR:Naming an artist in a prompt isn’t style control. T2I models tend to fall back oncanonical shortcuts— the recurring motifs, stock palettes, and over-represented period signatures they associate with that name — and quietly overwrite the scene the user actually asked for. This paper introducesAtelier, which plans an explicit control statebeforegeneration instead of hoping the backend infers intent.
What Atelier does
- Decomposes an underspecified request (“a quiet subway platform in the style of Van Gogh”) into an explicit control state:scene anchors,preserve vs. transformdecisions,style-regime hypotheses,role-bound artist evidence, andanti-shortcut constraints.
- Grounds that state with artist-level knowledge pluslocal patchreferences — and unlike generic RAG prompting, binds each patch to a specific scene role rather than dumping it in as an undifferentiated style exemplar.
- Compilesbackend-awaregeneration plans, then iteratively refines candidates with global + local authenticity critics.
ArtIntentBench A companion benchmark that trades artist breadth for supervision depth — two deliberately contrastive artists (Van GoghandQi Baishi, different media and cultural traditions) across four tasks: artwork re-rendering, period-controlled generation (Paris / Arles / Saint-Rémy / Auvers), historically unseen subjects, and Qi Baishi re-rendering, plus shortcut auditing and human preference eval. The argument for going deep: a wide benchmark of many artist names with shallow labels would end up measuring generic style association, not artist-grounded control.
Results Across both open-weight and closed-source generators, Atelier improves artist-level style fidelity, preserves source structure better, and substantially cuts shortcut substitution relative to prompt-engineered, retrieval-augmented, and general-purpose agent baselines.
**Why it’s interesting:**the framing that artist-grounded generation is bottleneckedupstreamof the diffusion model — in inferring explicit, evidence-grounded controls — is a nice counterpoint to the “just scale the generator” default. The shortcut-auditing setup also seems reusable well beyond these two artists.
Curious how the control state holds up on artists with less catalogued, more diffuse oeuvres, and whether the period-regime hypotheses transfer to movements rather than individuals.
Similar Articles
@AdinaYakup: Paper:
A new creator-centric benchmark for text-to-image generation, Qwen-Image-Bench, evaluates models on real-world fidelity and creative generation using a hierarchical taxonomy of 56 verifiable facets scored by a unified judge model.
STAR: SpatioTemporal Adaptive Reward Allocation for Text-to-Image RL Post-Training
This paper introduces STAR, a method for spatiotemporally adaptive reward allocation in RL post-training for text-to-image diffusion models, improving compositional alignment and text rendering by focusing policy updates on relevant latent regions.
ICML 2026 spotlight: Universal Aesthetic Alignment Narrows Artistic Expression \[R]
This ICML 2026 spotlight position paper identifies a failure mode in image-generation alignment where aesthetic preference optimization overrides explicit user intent, terming it 'reversed alignment' and testing on anti-aesthetic prompts.
ToolArtist: Tool-Using Unified Multimodal Models for Agentic Image Generation
Introduces ToolArtist, a fully agentic image generation model built from a unified multimodal model, using SFT and reinforcement learning (RAD-GRPO) to dynamically orchestrate reasoning, tool use, and image generation.
From Plans to Pixels: Learning to Plan and Orchestrate for Open-Ended Image Editing
An experiential framework for long-horizon image editing that couples planning with reward-driven execution to improve coherence and reliability of complex multi-step edits.