The models powering ChemIllusion

ChemIllusion blends transformer and diffusion technology, matching each model to the task it handles best.

ChemIllusion blends transformer and diffusion technology to craft chemistry-themed imagery. Each model is paired with tasks that suit its underlying architecture.

  • GPT‑Image‑1 – a diffusion model with cross-attention layers tied to a GPT-4.1 prompt encoder. It captures intricate molecular arrangements but is slower and costlier than lighter options.

  • GPT‑Image‑1.5 – an improved version of GPT‑Image‑1 with enhanced prompt understanding and better handling of complex chemistry scenes. Offers a balance between quality and performance.

  • GPT‑Image‑1 Mini – a lighter, faster variant of GPT‑Image‑1 optimized for quick iterations and lower-cost generation while maintaining good quality for simpler scenes.

  • DALL‑E 2 & DALL‑E 3 – CLIP-guided diffusion systems optimized for quick and dirty results. They react quickly but limit prompt length and output size.

  • Gemini 3 Pro Image Preview (Nano‑Banana) – Google's TPU-powered text-to-image diffusion model excels at photorealistic scenes and complex backgrounds. It is one of our best soft edit tools.

  • FLUX 2 Pro & Max – latent diffusion inpainting models from Black Forest Labs. Max yields high detail while Pro trades fidelity for speed and lower cost.

  • FLUX‑1 Fill – a lighter latent diffusion variant tailored for filling small regions; it's fast yet sometimes misses fine bond geometry.

  • Bria Reimagine – a copyright-safe diffusion model trained on licensed materials. Ideal for commercial use and publications where copyright certainty is important.

  • GPT‑5 family – mixture‑of‑experts language models used for analyzing images, suggesting prompts, or other text. They don't generate images but coordinate the rest of the system.

By matching each technology to the task it handles best, ChemIllusion balances quality, speed and cost across the platform.