Why do AI-generated visuals still look fake even after iterating on the same model?

Iterating on the same model resamples the same distribution — so the same plastic skin, uniform lighting, averaged faces, and over-clean backgrounds keep coming back. The fix is structural: render the shot across multiple models in parallel, pick the one that broke the pattern, and lock real reference inputs that anchor the output away from generic AI defaults.

Stop iterating in place. When a shot keeps looking fake on one model, the next generation on that same model is statistically going to land in the same neighborhood — same texture priors, same lighting averages, same symmetrical face library. Hridaye, invideo's creative director, put it directly: "When the jewellery ad looks too 'AI' - stop iterating on one model. Ask the Agent to render the shot across each, then decide."

invideo is an agentic video creation tool with every current image and video model available inside it — Recraft, Nano Banana, GPT-Image-2 on the image side; Veo, Kling, Seedance 2.0 on the video side — and the invideo agent routes each shot to the right one, so you never leave the platform to swap models.

Here is what to actually do, in order:

Render the shot across multiple models in one pass. Ask the invideo agent to generate the identical prompt across the available models simultaneously and compare side by side in the same chat. In one documented jewelry production, six models were tested against the same shot — the winning combination was building the base image in GPT-Image-2 for aesthetic, then passing it through Nano Banana to lock the exact product. Different models carry different texture vocabularies and lighting priors; the comparison is what surfaces the one that doesn't read as AI.

Lock real reference images, not just text. Generic prompts produce generic props and faces. When a Tiffany workshop was generated without real reference, the agent "invented generic-looking tools"; with real workshop references in context, the output matched a real Tiffany workshop. Upload product angles, real-environment photos, and (for products) a hand-holding-the-product shot so the model learns true scale. Decompose references by attribute — pull pose from one image, lighting from another, framing from a third — rather than feeding one composite and hoping.

Validate with a single probe shot before batching. Generate one full-frame shot containing character, garment, set, skin, and pose. If that one holds, the system is ready to scale; if it doesn't, fix the inputs, not the iteration count. Batching before the probe propagates the fake-looking failure across the entire run and burns credits.

Direct in craft language, not similarity language. "Make it more realistic" is the prompt that keeps producing plastic. Specify the depth structure (foreground props, subject plane, midground objects, backdrop), the light ("single hard warm key, side-raked, rim-lit skin, environment falling into shadow"), and per-shot pose at the micro level (weight distribution, hand position, jaw tension, gaze angle). One documented editorial campaign produced 40 stills and 30 motion clips at ~$150 total once those craft-specific rules were baked into the agent's context — and the campaign deliberately leaned into visible set construction rather than chasing photoreal smoothness.

Watch for over-literal prompts. "Stillness" produced models frozen stiff in one production until the rule was clarified to "one slow gesture, not zero." Fake-looking output is often the model obeying you too literally; loosen the rule with one micro-gesture or one imperfection and the uncanny edge drops away.

There's a credit angle too: re-rolling the same model 10 times to escape the same failure mode is the expensive way to stay stuck. Across documented productions, clip rejection runs ~85% even in well-tuned workflows — a multi-model comparison up front converts that waste into signal, because each rejected model tells you something the next one needs to do differently.

These are the levers that actually move AI visuals off the "fake" plateau — what works depends on your shot, but the order (multi-model probe → real references → craft-level direction) holds.

Watch some of these to see what works for you:

When the jewellery ad looks too 'AI' - stop iterating on one model. Ask the Agent to render the shot across each, then decide.

— Hridaye, invideo's creative director