| README.md | ||
A Full Image-to-Video Desk Loop Before the Next Meeting
Coffee-break length intro, then a full desk loop you can run before the next meeting.
Most creators treat a free AI image generator like a slot machine: type a long prompt, hope for a miracle, then rewrite everything when the next frame fails. That habit burns time and makes motion work almost impossible, because there is never a stable still to animate. I rebuilt the desk around a smaller rule set and kept the whole loop inside Yeri AI so prompts, references, and short clips stay in one project.
Start with four locked slots
Before generating anything, write four slots on a sticky note and refuse to rewrite three of them mid-batch:
- Subject — product, portrait, packaging, workspace object
- Surface — oak desk, marble, kraft paper, painted wall
- Light — soft window light, clean highlight, overcast skylight
- Style — editorial photo, clean product shot, flat illustration
Only the subject is allowed to change between runs. Light and style stay frozen for the whole session. That single constraint is what makes a free AI image generator feel repeatable instead of random.
Generate for direction, not polish
The first round is a scout. A prompt like "ceramic bottle on pale oak, soft window light, editorial photo" is enough. Generate a handful of variations and score them on three binary checks: is the composition readable, is the subject complete, and does the colour fit the brand board. If none pass, change one variable and run again. Rewriting the entire prompt every time is the fastest way to lose a good direction.
Comparing models helps at this stage. Seedream tends to give crisp product lighting. Nano Banana is quick for iterations. GPT Image is stronger when packaging text or layout matters. Because they sit in the same project, the comparison takes minutes instead of an afternoon of exports and re-uploads.
Clean after composition locks
Background remover, upscaler, and light templates come last. There is no point polishing a frame you will discard. I wait until one still clearly wins, then clean it once. Upscaling a throwaway frame wastes credits and attention.
Motion as a six-second gate
When the still is stable, push it into image to video. The first clip is not a finished ad. It answers two questions only: does camera movement distract from the subject, and does the pacing hold for six to ten seconds. Kling and Seedance disagree often. Keep the calmer take. If the subject drifts, go back to the still instead of fighting the timeline.
That short gate saves the longest failures. A weak still will not become a strong fifteen-second clip no matter how many motion settings you twist.
Archive the recipe
Save the winning prompt block, the model name, the reference still, and one sentence about why it worked. Next session starts from a known-good recipe instead of a blank box. Over a few weeks this archive becomes more valuable than any single generated image.
What I refuse to do
- I do not hop between five AI tabs mid-batch.
- I do not change light and style words at the same time as the subject.
- I do not upscale before composition is locked.
- I do not start a long video render before a six-second test passes.
Closing checklist
- Lock light and style
- Change one variable at a time
- Edit after composition
- Test short motion
- Archive the recipe in the free AI image and video workspace
Host note: drafted for hedgedoc.private.coffee on 2026-10-05 with a longer desk-routine focus so the page reads like working notes rather than a thin directory blurb.