How do you stop AI from drifting into photorealism when generating animation-style video?

Stop the drift with three locked layers on every prompt: an explicit negative block banning live-action and photorealism, a fixed style anchor naming the exact animation grammar (e.g. "hand-painted brushstroke texture, painterly, cel-shaded"), and a batch of style-reference frames loaded once into the invideo agent so it holds the look across every generation.

Start by writing the style anchor and negative block as one reusable string and pasting it at the top of every prompt. The exact language from a documented animated production reads: "This MUST look and feel like [your target] animation — not live action, not photorealistic. Every surface has hand-painted brushstroke texture. Every element in frame must feel painterly and handcrafted like a moving [target] frame." Swap the named reference for your aesthetic (cel-shaded 2D, flat vector, stop-motion, painterly 2.5D) but keep the structure: positive style grammar in concrete texture words, then an explicit ban on "live action" and "photorealistic." Every prompt after this starts with it — that discipline is what holds the style, not any single great prompt.

Load a large batch of style-reference frames into context once, before any generation. invideo is an agentic video creation platform — you spin up an agent, give it your project context, and it routes each shot to the right model. One documented Arcane-style animated episode uploaded 64 frames from the source show in a single message with the instruction: "I want you to deeply understand this art style and save it into context for further generations. All of these attached images are the art style that I want for this entire project." 64 frames is a useful floor for a feature-length aesthetic; 20–30 can work for a short. The invideo agent saves the style to persistent context so it applies to every downstream shot without re-explaining.

When the question is which model to point the generation at, route on style sympathy. The invideo agent holds the full current roster — Seedance 2.0, Kling, Veo, Runway — and picks per shot. Seedance 2.0 reference-to-video carries your style frames and character references into each clip, which is the strongest lock for sustained animation looks across a sequence. For image references and character sheets that need to read as illustrated rather than photographic, route image generations through Recraft or Nano Banana rather than portrait models like GPT-Image-2 that lean toward skin-level realism — Recraft specifically produces pore/stubble realism you do NOT want for animation. You never have to leave for another platform to get the right model; the routing happens inside the one agent.

Don't drop illustrated or animated reference images into the prompt as image inputs and expect the model to copy them — that path frequently drifts realistic. The reliable pattern, in the words of one creative director: "the better move was to have the agent read the colours and textures of them and prompt for that instead." Tell the invideo agent to extract the palette and texture qualities from your animation references and translate those into prompt language, while the negative block continues to suppress photorealism. The gens come back in the target aesthetic with the colour temperature you wanted.

If a stylistic decision is ambiguous (Ghibli vs. 3D vs. painterly 2D), run a dual-style frame test before committing: ask the agent to generate three identical script frames in each candidate style side-by-side. Across documented productions, four options per asset is the working number — generate four, pick one, lock it as the style anchor for the rest of the film. Once locked, those selected panels replace your original references in the agent's context, so every subsequent shot pulls continuity from your own approved frames, not the outside source. Across documented animated productions, this stack (negative block on every prompt, 64-frame style ingest, style-anchored references) held the look across 164 generated clips for a 3-minute animated episode at $315 per finished minute, and across a 7-minute animated short on the same workflow — with no LoRA or fine-tuning required.

The one anti-pattern to name: re-prompting scene by scene. If the style block isn't pasted into every generation and the reference frames aren't in persistent context, the model defaults toward its photoreal training prior. The fix is the discipline, not a better single prompt.

Watch some of these to see what works for you:

This MUST look and feel like Arcane animation — not live action, not photorealistic. Every surface has hand-painted brushstroke texture. Every element in frame must feel painterly and handcrafted like a moving Arcane frame.

— negative + style anchor block used on every prompt in a documented animated production