[Paper] DraCo: Draft as CoT for Text-to-Image Preview and Rare Concept Generation

Source: arXiv - 2512.05112v1

Overview

The paper introduces DraCo (Draft-as‑CoT), a new way for multimodal large language models to generate images from text. Instead of relying solely on textual “chain‑of‑thought” planning, DraCo first creates a low‑resolution draft as a visual sketch, then uses the model’s reasoning abilities to spot and fix mismatches before producing the final high‑resolution image. This interleaved text‑and‑image reasoning dramatically improves the fidelity of generated pictures, especially for rare or complex concepts.

Key Contributions

• Draft‑as‑CoT paradigm: Treats a low‑resolution draft image as an explicit step in the chain‑of‑thought, enabling concrete visual planning and verification.
• DraCo‑240K dataset: Curated 240 K training examples covering three atomic skills – general correction, instance manipulation, and layout reorganization – to teach the model how to refine drafts.
• DraCo‑CFG: A specialized classifier‑free guidance technique that harmonizes interleaved textual and visual reasoning during generation.
• Significant performance gains: Improves benchmark scores by +8 % on GenEval, +0.91 on Imagine‑Bench, and +3 % on GenEval++ compared with standard text‑only CoT or direct generation.
• Rare concept handling: Demonstrates robust generation of uncommon attribute combinations that typically fail for existing models.

Methodology

1. Prompt → Draft

   • The model receives a natural‑language prompt and first generates a low‑resolution draft image (e.g., 64×64).
   • This draft acts as a visual “thought” that captures coarse layout, object presence, and rough attributes.

2. Verification & Error Detection

   • Using its internal multimodal understanding, the model compares the draft against the original prompt.
   • It identifies semantic gaps (e.g., missing objects, wrong colors, misplaced layout).

3. Selective Refinement

   • The model decides which parts need correction and applies targeted edits (instance addition/removal, attribute adjustment, layout shift).
   • A super‑resolution module upsamples the corrected draft to the final resolution (e.g., 512×512).

4. Training with DraCo‑240K

   • The dataset provides paired examples of prompts, drafts, and corrected high‑res images, annotated for the three atomic capabilities.
   • Losses combine standard diffusion objectives with auxiliary supervision for correction decisions.

5. DraCo‑CFG Guidance

   • Extends classifier‑free guidance to operate on both the textual and visual branches simultaneously, ensuring the draft and final image stay aligned with the prompt throughout the diffusion process.

Results & Findings

• Qualitative gains: Visual examples show sharper object boundaries, correct rare attribute pairings (e.g., “a teal‑striped zebra”), and more faithful spatial arrangements.
• Ablation studies: Removing the draft step drops performance by ~5 % on GenEval, confirming the draft’s role as a crucial planning scaffold.
• Error analysis: Remaining failures are mostly due to extreme prompt ambiguity rather than model incapability.

Practical Implications

• Rapid prototyping for designers: Developers can obtain an instant low‑res preview, iterate on the prompt, and let the model auto‑refine, cutting down on trial‑and‑error cycles.
• Content creation pipelines: Integration into asset‑generation tools (games, AR/VR, advertising) where rare or custom concepts are common.
• Improved safety & controllability: The verification step can be extended to enforce policy constraints (e.g., no disallowed objects) before upscaling.
• Reduced compute waste: By catching major mismatches early at low resolution, the system avoids expensive high‑res diffusion on obviously wrong drafts.

Limitations & Future Work

• Draft quality ceiling: Very low‑resolution drafts sometimes miss fine‑grained details, limiting the model’s ability to correct subtle errors.
• Scalability to ultra‑high resolutions: Super‑resolution still relies on standard diffusion upscalers; integrating dedicated upscaling networks could improve fidelity.
• Prompt ambiguity handling: The current verification assumes a well‑specified prompt; future work could incorporate interactive clarification loops with users.
• Dataset bias: DraCo‑240K, while diverse, may under‑represent certain domains (e.g., medical imaging), suggesting the need for domain‑specific fine‑tuning.

DraCo opens a new avenue where visual drafts become an integral part of a model’s reasoning chain, bridging the gap between abstract textual planning and concrete image synthesis. For developers building next‑generation generative tools, this approach promises more reliable, controllable, and creative outputs.

Authors

• Dongzhi Jiang
• Renrui Zhang
• Haodong Li
• Zhuofan Zong
• Ziyu Guo
• Jun He
• Claire Guo
• Junyan Ye
• Rongyao Fang
• Weijia Li
• Rui Liu
• Hongsheng Li

Paper Information

• arXiv ID: 2512.05112v1
• Categories: cs.CV, cs.AI, cs.CL, cs.LG
• Published: December 4, 2025
• PDF: Download PDF