If AI image generation is supposed to be the Photoshop killer why are chemists still spending hours in photoshop?
AI tools promise to replace manual editing, yet chemists still rely on Photoshop for fine-tuning and scientific accuracy.
AI image generation tools often produce aesthetically pleasing outputs, yet they can stray from what a prompt demands. These models are trained on billions of images and captions, so they learn to match visual patterns with general language concepts, but not to reason about every detail. When a chemist requests “a benzene ring with an attached methyl group,” the model might deliver something chemically inspired but incorrect—text become garbled, atoms miscounted, or spatial relationships distorted. The engine is optimizing for plausibility, not precision. In other words, these systems are great at style but loose on exact instructions.
Prompt engineering and context engineering help tighten this gap. Prompt engineering clarifies style, perspective, and fine details (“top‑down schematic, white background, carbon atoms labeled”). Context engineering supplements the prompt with structured data—think SMILES strings, known 3D coordinates, or a short description of the molecule. When the generator sees both the request and context, it can map the text to a cleaner latent representation and adhere more closely to the user’s intent.
ChemIllusion exploits both strategies. For each request, it funnels the user’s description and any attached chemical data into engineered prompts tailored to the model being used. It keeps a model catalog: certain diffusion models handle cartoon‑style atoms, others excel at photorealistic lab equipment, and still others serve text overlays or vector icons. ChemIllusion evaluates which model has the best track record on a given input (using metadata on molecule size, style, and background) and routes the job accordingly. By pairing the right prompt structure with the right engine, it avoids the common prompt‑adherence pitfalls that force chemists back to Photoshop.
The “Adjust Scene” feature then brings surgical control. Instead of rewriting the whole prompt, a chemist can click a button to say, “Add a 50 mL labeled flask on the left and ‘Acetone’ text in the center.” Adjust Scene interprets this at two levels:
Text parsing – it extracts object names, sizes, and text labels.
Spatial placement – it converts phrases like “on the left” or “in the center” into explicit coordinates and masks.
Using those masks, ChemIllusion performs targeted inpainting. Objects and labels are placed precisely where requested, preserving the rest of the image. The heavy lifting—clean borders, correct fonts, lighting—happens automatically. The result is a scene that looks finished without ever opening Photoshop.