CRAFT: Continuous Reasoning and Agentic Feedback Tuning
for Multimodal Text-to-Image Generation
Hugging Face Release 
Parti-Prompts: 1.6K+ long, complex prompts testing spatial reasoning, composition, and text rendering.
Same neural network with thinking mode and "off" and "ON"
Abstract
CRAFT brings explicit “thinking” into image generation. Instead of generating images in a single shot, it verifies visual constraints, fixes only what fails, and stops when all constraints are satisfied — improving reliability without retraining or scaling models.
CRAFT (Continuous Reasoning and Agentic Feedback Tuning) is a training-free, model-agnostic framework for text-to-image generation and image editing. It converts a user prompt into dependency-structured visual constraints, verifies generated images using a vision–language model, and applies targeted prompt updates only where constraints fail.
CRAFT runs entirely at inference time and works on top of existing image generators, requiring no retraining or architectural changes. With minimal overhead, it consistently improves compositional accuracy, text rendering, and preference-based quality across multiple model families — allowing small and inexpensive models to approach the performance of much larger systems.
Cost-quality trade-off at 2048×2048 resolution. CRAFT cost includes two generations plus $0.00128 reasoning overhead (VLM + LLM). Cost multiplier shown in parentheses.
Methodology
Methodology CRAFT operates as a small multimodal orchestrator: given a user prompt (1), it decomposes the intended scene into explicit visual questions (2), evaluates generated images (4) with a VLM through visual-question answering (5), and rewrites the prompt (8) with an LLM to fix the detected inconsistencies. At each iteration, the system alternates between generation (4), reasoning via VQ responses (5, 7), and correction of the current prompt (3, 8), using deterministic visual questions (DVQ) as a stable signal for constraint satisfaction. A comparator (6) checks whether all constraints are satisfied; the loop continues until they are met or a small iteration budget is exhausted. CRAFT is fully model-agnostic and works with any T2I backbone (diffusion, autoregressive, or API-based). It requires no training, introduces minimal overhead, and reliably improves compositional accuracy, text rendering, spatial relations, and artifact detection across a wide range of models. In essence, CRAFT acts as an embedded agentic orchestrator for image generation: a lightweight reasoning supervisor that ensures the final output faithfully matches the user’s intent (loop 1→2→3→4→5→6→7→8).
Results
CRAFT improves prompt–image alignment on DSG1K and Parti-Prompts across all evaluated T2I models. It consistently increases VQA and DSG scores, with the largest gains on weaker backbones, and is strongly preferred in automatic side-by-side evaluations. The method is training-free, model-agnostic, and effective with only a single reasoning iteration.
DSG1K Evaluation
DSG1K: ~1K prompts evaluated via structured yes/no visual questions for fine-grained compositional correctness.
Parti-Prompts
Parti-Prompts: 1.6K+ long, complex prompts testing spatial reasoning, composition, and text rendering.
Human Evaluation
Human evaluation further confirms these trends. Across representative architectures — Z-Image-Turbo, FLUX-2 Pro, FLUX-Dev, and FLUX-Schnell — annotators prefer CRAFT over the baseline by wide margins
BibTeX
@article{craft,
title={CRAFT: Continuous Reasoning and Agentic Feedback Tuning or Multimodal Text-to-Image Generation},
author={V. Kovalev, A. Kuvshinov, A. Buzovkin, D. Pokidov, D. Timonin},
journal={arXiv preprint arXiv:2512.20362},
year={2025}
}