Write shots the way cinematographers do
The quality of a generated board depends on the vocabulary in your prompts. Use real shot grammar: "wide establishing shot", "over-the-shoulder", "low-angle close-up", "dutch angle". Specify lens feel ("long lens compression", "wide 24mm distortion") and lighting continuity ("overcast noon", "single practical lamp"). The model understands cinematography language because it learned from cinema.
Note the camera motion for each panel in your annotations — dolly-in, whip-pan, crane-up — using the same terms as Nidhogg's camera presets. When you later animate a frame, the annotation becomes the preset choice.
The continuity trick: fixed descriptor tokens
The classic AI storyboard failure is your protagonist changing face, wardrobe, and age between panels. The fix is mechanical: write one exact descriptor string per character and location — "a wiry man in his 50s, grey stubble, worn brown leather jacket" — and paste it verbatim into every frame's prompt. Identical words in, consistent character out. Keep these tokens in a document beside your shot list and never freestyle them.
From board to animatic
Iterate frames on flux-schnell — it's fast and cheap enough to try four compositions per shot — and apply CINEMATIC TEAL across the batch so the board reads as one film rather than twenty disconnected images. Then take your five or six pivotal frames and run them through a video model with the camera move you annotated. Cutting those clips to a temp track tells you more about pacing in ten minutes than a week of imagining it.