Academy
Guides for getting better generations
Short, practical write-ups, not fluff. Video walkthroughs are on the roadmap; for now these are the same notes we'd give a teammate.
Video lesson (coming soon) · 400×225
Writing prompts that actually hold up
- •Be literal about camera, subject, action and setting, in that order. Most models weight earlier tokens more heavily.
- •State what should move and what should stay still for video models; ambiguity here is the #1 cause of warped limbs and flickering backgrounds.
- •Negative prompts help image models more than video ones. For video, describing the desired result usually beats describing what to avoid.
Video lesson (coming soon) · 400×225
Picking the right model for the job
- •Flux Schnell / fast tiers are for iteration: cheap, quick, rough. Move to a pro-tier model only once the composition is locked.
- •For talking-head or lip-synced content, prioritize models explicitly trained for it (check the model hint in the selector). General video models handle faces poorly.
- •Chat models: smaller/cheaper models are fine for captions and short copy; save the frontier models for anything requiring multi-step reasoning.
Video lesson (coming soon) · 400×225
Keeping a character consistent across shots
- •Reuse the exact same subject description string across every generation in a sequence. Small wording drift compounds into visibly different faces.
- •Where a model supports a reference/seed image, always pass one once you have a shot you like. Regenerating from text alone resets everything.
- •Lock your aspect ratio and lighting description early; changing them mid-project is the most common reason a sequence looks disjointed.
Video lesson (coming soon) · 400×225
Formats for short-form video (Reels / Shorts / TikTok)
- •9:16 vertical, 15-60s. Say the platform's target duration explicitly in the prompt: models default to horizontal, cinematic pacing otherwise.
- •Hook in the first 1-2 seconds: describe the opening frame as the most visually striking moment, not a slow build-up.
- •Generate B-roll separately from your hook shot. Trying to get both in one generation usually dilutes both.
Video lesson (coming soon) · 400×225
Budgeting AI Credits across a project
- •Draft in the cheapest tier for every model family first; only your final 2-3 picks should ever hit a pro-tier model.
- •Video is priced per second. Trimming a test generation from 10s to 4s cuts cost proportionally with no quality loss for composition checks.
- •Check the Dashboard's transaction log after a session: it shows exactly which provider and model billed each generation, useful for spotting an expensive habit early.