How do you keep a two-photo AI video series visually consistent across clips?

I'm building a short-form video workflow where each clip is generated from a pair of still photos rather than filmed footage, and I've run into a consistency problem that I suspect other writers here have hit with any serial visual project.

**The setup**

Each clip takes two source photos, applies a preset motion direction, holds the camera timing steady, and outputs at 4:3 and 720p with a generated audio bed. The output is a synchronized two-person scene, which is exactly the format a lot of short-stay and venue clients ask for.

**Where the consistency breaks down**

- The generated motion is not identical between clips, so a series has a visible "jump" every time the model reinterprets the same motion preset.
- Camera distance and eye level drift between the two source photos, and the drift reads as a slide rather than a performance.
- The generated audio is regenerated per clip, so the bed does not sit continuously across a multi-clip sequence.

**What I've settled on so far**

1. Fix the template length and output settings for the whole series and never vary them mid-project.
2. Match camera distance and eye level across both source photos before generating anything.
3. Keep the background deliberately simple, because every extra object is something the motion model may warp between frames.
4. Review every clip in four passes before it goes anywhere near a client.

**What I'm still working out**

Does anyone have a better method for keeping the motion signature stable across a series? I've been comparing outputs against a reference implementation, and this [hotel lobby ai generator](https://hotellobbyaivideo.online/) is the one I've been using as a baseline because it keeps the template length and motion presets fixed, which makes the differences easier to isolate. But I'm genuinely curious whether a serial consistency workflow exists that I'm missing, particularly for writers who produce regular visual serials rather than one-off pieces.

Curious how others handle continuity when the underlying generator introduces run-to-run variation.