Why do title cards and text overlays look inconsistent across AI-generated video projects?

Title cards look inconsistent because each generation is a stateless event: the model holds no memory of the previous card's typeface, weight, kerning, or color temperature, and cards produced in separate sessions or by separate sub-agents pull from no shared style reference. Most image and video models also render text unreliably, adding distortion on top of drift.

Three causes explain the mismatch, and each has a direct fix.

Every title card generation is independent. Generative models don't carry typography forward — a card rendered today and a card rendered tomorrow are two unrelated samples, so font choice, letter spacing, color temperature, and layout all re-roll unless you pin them. This is the same context-loss problem that causes character drift, applied to type: tools without persistent memory lose your visual language between clips, costing roughly 20 minutes per session in re-described setup before you even reach the drift itself.

Cards produced in different sessions or by different agents don't match. This is documented directly: title cards generated by different agents or sessions inside a project will visually mismatch, which is why some productions generate all title cards in post-production — building them in your editor (DaVinci Resolve or any NLE) guarantees pixel-identical type across the whole film. If you want the cards generated rather than built, the fix is the mirror of the cause: produce every card in one session, from one sub-agent, against one locked reference. invideo is an agentic video creation tool where project context persists, so a practical setup is a dedicated title-card sub-agent — one documented multi-agent workflow assigned a single agent to B-roll and title cards specifically so all cards pulled from the same shared context brief. Lock one approved card as a reference image in project context and attach it to every subsequent text-overlay request; the invideo agent then holds that style across the project instead of re-inventing it per card. Title cards are a normal agent deliverable — one 45-second production shipped 26 video clips plus 4 text cards as a single asset folder — so routing them through one context-holding agent costs nothing extra.

Text rendering quality is model-specific. Most image models distort or garble type: Nano Banana Pro is strong for compositional imagery but produces garbled text on graphics and logos, while GPT-Image-2 renders text without distortion — the working rule from documented productions is to route any image containing text to GPT-Image-2 and keep Nano Banana Pro for non-text frames. Inside invideo you don't have to leave the project to switch: the invideo agent routes each generation to the right model, so tell it explicitly that all title cards and text overlays go through GPT-Image-2.

In short: pin the typography with one locked reference card, generate all cards in one session or one sub-agent, route text imagery to a text-reliable model — or skip generation entirely and set the type in your edit, where consistency is free.

Watch some of these to see what works for you:

Every single one of these tools has amnesia. You spend 20 minutes setting up your character, your world, your visual language, generate a clip, it looks great, then you move to the next scene, and the tool has forgotten everything.

— a creator documenting AI video production workflows