Phantra AI: Series And Continuity
Maintain visual continuity across Phantra episodes by reusing the same project, references, style, length, frame, and model.
Series and continuity
The studio renders each scene, and each shot in it, on its own; no render
remembers the last one. What makes two scenes read as one series is what you
carry across on purpose: the same photos in the same order with the same
roles, the same style block word for word, the same length, frame and model.
Setting a series up and writing its episodes spends nothing; a render is
quote, the person's yes in chat, then approve (price-and-render).
Set up once
- One project.
projects; when none fits,create_projectwith atitle, abrief(who it is for, what every episode must always show and never show) and alook(the style block, in cinematography terms). Every scene in the series is saved with itsproject_id. - One cast.
elements: the characters, products and places the series uses. Decide the ref ORDER once (the anchor first: the face before the room) and therolesonce ({ref: role}). Write both down in the chat so every episode uses the same list. - One style block: two or three sentences (palette, light, lens and framing
height, depth of field, finish). It is the project's
look; paste it into every scene unchanged. - Fixed defaults:
set_preferenceswithsecondsandvideo_model, sowrite_sceneandquotecalled with nothing else still match the series; the 9:16 shape is the studio's own frame, not a preference.preferencesreads them back; a value the catalogue lacks is refused.
Each new episode
projectwith the id: the brief, the look, the scenes so far and the references they used. Thensceneon the last episode'sscene_idfor its full prompt (projectshows only the start of each).duplicate_scenefrom the last episode'sscene_id, passing only the newprompt(andsecondswhen it changes).refs,rolesandproject_idare copied, so the order and the roles hold. Its stills and clips are not copied: the copy is a fresh scene.- In the new prompt change only the beats. Keep the opening line and the style block word for word, keep the sound line's shape, keep the avoid list, and name the same photos the same way.
- Show it to the person, then
quoteit: the keyframes (the credits foreachstill) and the clip, one render per shot with its credits and a token, against the balance; with a series default set,preferencesays which model it priced. Say the price in credits and WAIT for the person's yes in chat. - Only on that yes:
approvewith the quote's tokens and the same renderer choice (provider, model, frame, duration), thenjobuntil done andscenefor the clips. Never approve on the person's behalf, never twice;insufficient_credits(needs, available) means stop and say credits are added in the studio (price-and-render).
Episode 2 from scene 418, with episode 1's opening line and style block pasted in unchanged (replace the beats):
duplicate_scene(scene_id=418, prompt="Ultra-realistic grounded deadpan comedy
in 9:16. The attached photo of Sam is his exact face and clothing. Style: soft
daylight from one window, true-to-life colour with gentle contrast, a 35mm
frame at chest height, shallow depth when close, practical light only.
(0-5s) Sam opens the same dented tin on the same kitchen counter; inside, a
folded map.
(5-10s) Hard cut to the hallway, the framing of episode 1: Sam reads the map,
and the front door is already open. End abruptly.
No background music. Only diegetic sound: paper, the hinge.
Avoid: plastic AI sheen, CGI-looking skin or fabric, cartoon reactions,
dramatic slow motion, named films, brands or actors.")
Check continuity after the renders
scene on each episode lists the clips rendered so far. Read them against
the series rules: the same frame and length, the anchor's face or product in
every episode, the same light and palette. A clip that drifted is not edited
here: say what drifted (a changed style sentence, a ref in a different order,
a different model or frame), fix the prompt or the refs, and duplicate_scene
again as a new scene, quoted and approved on a yes like any episode. Over
time what_works says which model and preset the person kept most, which is
the one to hold the series to.
Do not
- Never reorder the refs between episodes: the first ref is the anchor, and a new first ref is a new look.
- Never rewrite the style block to improve one episode; change it for the
whole series: the person changes the project's look in the studio (no tool
here does), the new block goes into every episode written after it, and an
earlier episode that must match is made again with
duplicate_scene, never rewritten in place. - Never move one episode to another model or frame to save credits without
saying the look will shift; price the draft frame instead, where the series'
model has one (
video_modelssays which do):set_preferenceswithdraft_firsttrue, thenquotewith no renderer named, andapprovenaming the model and frame the quote answered. - Never promise two renders will match exactly. What you can promise is the same photos, words, length, frame and model.
- Downloads
- Security audit
- Pass
- Last updated
- 2w ago
- Current version
- v1.0.0
- License
- MIT-0
