An image generator that finally learned to spell

What it does PosterCraft is an ICLR 2026 research project that generates aesthetic posters from text prompts. It builds on the FLUX.1-dev diffusion model and swaps in custom transformer weights trained for poster composition, typography, and style harmony. The target output is event promos, movie posters, and product ads where the text is legible and the layout looks intentional.

The interesting bit Training is a four-stage assembly line instead of a single fine-tuning pass: text rendering, poster-specific aesthetic tuning, reinforcement learning against aesthetic-text preferences, and finally joint vision-language feedback using Qwen3-8B. Looping in an LLM to iteratively critique generated layouts is a notable departure from standard diffusion fine-tuning.

Key highlights

• Project-reported benchmarks place it near or above closed models like Gemini 2.0 Flash Gen and Ideogram-v2 on text accuracy.
• Ships two checkpoints: PosterCraft-v1_RL (preference-optimized) and PosterCraft-v1_Reflect (vision-language refined).
• Released training datasets including Text-Render-2M and HQ-Poster-100K on HuggingFace.
• Community ComfyUI integration exists alongside an official Gradio demo and HuggingFace Space.
• Inference requires the FLUX.1-dev pipeline plus Qwen3-8B; an offload script is provided for memory-constrained GPUs.

Caveats

• Benchmarks are author-reported without independent third-party evaluation.
• The full pipeline is resource-heavy; even offload mode needs both a large diffusion model and an 8B-parameter LLM.
• The project is brand-new (June 2025), so behavior beyond the curated gallery remains unclear.

Verdict Try it if you need automated marketing or event posters and are exhausted by AI-generated typos. Avoid it if you want a lightweight, single-checkpoint model or lack the hardware to run FLUX alongside an 8B language model.