If you have spent an evening with text-to-video tools, you know the loop. You write a prompt, wait for the clip, and something is off. Maybe the room is wrong, maybe the character is facing the wrong way, maybe the motion is fine but the framing cuts off the one thing you wanted to show. So you rewrite the prompt and spend another generation finding out whether it worked.
The problem is that a single text-to-video prompt asks the model to solve two different things at once: what the scene looks like, and how it moves. When the result is wrong, you often cannot tell which half failed. Debugging gets slow and expensive.
The fix I have settled on is boring and very effective. Split the job in two. Get the still image right first, then animate it.
Why splitting the problem helps
Software people will recognise this immediately. If a function does two things and produces a bad result, you separate the steps so you can test each one on its own. AI video works the same way.
Step one is composition: subject, setting, lighting, camera angle, framing. Step two is motion: what moves, how fast, and what the camera does. Once the first frame is locked, the video prompt only has to describe change, and every failure you see afterwards is a motion problem. That is a much smaller space to search.
There is also a cost argument. Image generations are quick and cheap compared with video, and some are free. Iterating on composition at image prices, then spending video credits once on a frame you already like, beats paying video prices to discover that the background was wrong.
Step one: draft the frame
For quick drafts, I use a free GPT Image 2.5 generator that runs without an account or a card. You type a prompt, pick an aspect ratio and download the result at 1K, which is plenty for a first frame that will become a short clip. The free page works from text only, so if you need to guide the image with reference pictures you will have to move to the full workspace, but for blocking out a scene it does the job.
A few habits make this step pay off:
Pick the aspect ratio now, not later. If the video is going to be vertical for a phone, generate the still at 9:16. Cropping a landscape frame afterwards wastes the framing you just spent time on, and asking a video model to extend a frame invites it to invent things.
Leave room for the motion. If a person is going to turn their head or walk forward, give them space in the frame to do it. A subject jammed against the edge has nowhere to go.
Keep the important details large and simple. Small objects, fine patterns and tiny text are the first things to wobble once a frame starts moving. If a logo or a label has to stay readable, it probably belongs in post-production, not in the generated frame.
Generate several versions and choose deliberately. Treat it like picking a thumbnail. You are looking for the frame that already tells the story before anything moves.
Step two: animate the winner
With a frame you like, move to an image-to-video generator and upload it as the starting image. Now the prompt can be short and specific, because the model does not have to imagine the scene. Describe only what changes: “steam rises from the cup, the camera pushes in slowly”, or “she turns toward the window and the curtains move slightly in the breeze”.
I start with the shortest clip and the lowest resolution available, check whether the motion reads correctly, and only then render the version I intend to keep. The generator I use shows the credit cost before each run, so the expensive render is always a conscious decision rather than a surprise.
Reading the failures
The real benefit of the two-step approach is that problems become easy to diagnose.
If the scene looks wrong in the video, it was wrong in the still. Go back to step one.
If the scene looks right but the motion is strange, the motion prompt is asking for too much. Cut it down to one action and one camera instruction.
If faces or objects drift as the clip goes on, the clip is probably too long for the amount of change you requested. Shorten it, or reduce the motion.
If the clip is fine but feels lifeless, the frame may be too static. Choose a still that implies a next moment: someone mid-step, a door half open, liquid about to pour.
When text-to-video is still the better choice
None of this means text-to-video is useless. For abstract scenes, quick mood tests and ideas where you genuinely do not care about the exact composition, going straight from text is faster. The two-step approach earns its keep when the frame matters: product shots, characters that need to look consistent across several clips, or anything a client is going to approve.
In those cases, five minutes spent on a free still is the cheapest debugging you will ever do.
