4.1 KiB
Discard Early, Polish Once, Then Export
Written after too many afternoons spent polishing frames that should have been discarded.
Most free AI image generator workflows fail for the same reason: every run rewrites the whole prompt. The next frame never matches the last one, so motion work collapses. This note is a desk checklist I actually use inside Yeri AI when I need stills that can become six-second clips without starting from zero every hour.
Lock four slots before you generate
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. When people complain that models are "inconsistent," they usually mean the prompt kept drifting across a dozen tiny edits.
Scout first, polish later
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. Keeping them in one project avoids an afternoon of exports and re-uploads between tabs that never share history.
Clean only after composition locks
Background remover, upscaler, and light templates come last. There is no point polishing a frame you will discard. Wait until one still clearly wins, then clean it once. Upscaling a throwaway frame wastes credits and attention. The same rule applies to face clean-up and text fix passes: they belong after selection, not before. Treat polish as a single deliberate pass, not a nervous habit.
Motion as a six-second gate
Once a still passes, treat motion as a short proof, not a film. Six seconds is enough to see whether the camera path respects the subject and whether the lighting still reads. If the clip fails, return to the still and change one axis — camera height, push speed, or secondary motion — instead of regenerating the entire scene. Tools like Kling or Seedance work better when the reference still is stable and the brand colour has already been locked.
A practical sequence inside the same desk:
- freeze the still that won the scout
- generate two short clips with different camera moves
- keep the one that preserves brand colour and product edges
- only then export for social or an ad board
Why a single workspace matters
Jumping between disconnected tools is where consistency dies. Prompt history, references, and clip trials need to live next to each other. That is why I keep the loop in one place rather than stitching five browser tabs. If you want a concrete starting point for the same workflow, open Yeri AI and build one project with the four locked slots above before you generate anything. The habit of locking slots first is more important than any single model choice.
Checklist you can hand to a junior
- Four slots written before the first generate
- Scout round scored with three binary checks
- One polish pass after the winner is chosen
- Motion capped at six seconds for the proof
- Only one variable changed when a clip fails
- Exports happen after the still and clip both pass
Follow the list in order and you will spend less time arguing with the model and more time shipping frames that can actually move. Instance note for notes.fsi.uni-tuebingen.de.