Video Generation

Video Generation (FLUX 3)

Trigger

Any request to generate, animate, or chain video clips, or to produce a short film for a contest/deadline.

Mandatory first step

Call bfl_flux3_prompting_guide before the first generation of any session. It costs no generation budget and carries the prompting rules.

Tool selection

Hard constraints (plan around these)

• One generation = one 5-20s clip at 24fps. A 20-40s short film = 2-4 chained segments, or one strong single take.
• video_continuation: input mp4 ≤50MB and ≤15s; the new segment also caps at 15s. Chain a second continuation for longer sequences.
• keyframes_to_video: when pinning keyframe_indices, set an explicit duration — "auto" can resolve shorter than your highest index and the job is rejected.
• Audio is generated by default; write "no music" in the prompt if unwanted.
• Generation takes MINUTES per clip and polling is normal. Budget iteration time: each retry is minutes, not seconds. For deadlines, finish with hours of buffer.
• Media fields accept local paths (auto-uploaded to temporary storage, deleted after) or URLs.
• A generate call returns a job id; nothing exists until bfl_flux3_get_result returns Ready and downloads the clip.

Companion craft skill

This skill owns MECHANICS only. Prompt-writing craft (consensus prompt anatomy, cinematography vocabulary, suspense-reveal beat sequencing, verbatim prompt library, failure-mode fixes) lives in flux3-video-directing (creative category) when present — load it for any "write/improve/storyboard a prompt" work. Built 2026-07-30 from two research digests (~/flux3-video-prompting-research.md, ~/cinematography-ai-video-digest.md) via: research fan-out (2 parallel delegate_task scrapers) → senna writes a spec file deciding ONE-skill-not-a-series → file-relay to creative profile for authoring. Reuse that pattern for building craft knowledge around any new tool.

Prompt shape

Plain prose, ordered: subject → distinguishing visual specifics → action → camera → lighting → environment → audio → style. Whatever you leave unspecified gets filled in for you — research anything with a real checkable appearance (grounding=true allows a short research pass).

Timed competition workflow (see references/contest-brief-research.md)

1. RESEARCH THE BRIEF BEFORE DISPATCHING ANYONE. Rules, submission format, tags, exact deadline with timezone, judging criteria. Never dispatch specialists on a rumored brief.
2. Convert deadline to a work-backwards plan: post by deadline minus a multi-hour buffer (iteration is slow).
3. Feasibility check against the constraints above BEFORE promising a film length.
4. Specialists in parallel (concept/storyboard, post copy) → synthesize one plan → user green-light → generate.
5. Deliver the exact submission artifact the rules require (e.g. file + caption + tags), not just the video.

Rate limits & multi-profile coordination (learned 2026-07-30)

• HARD LIMIT: one new generation per account every ~5 minutes, and only one job active at a time. Submitting early returns a cooldown error naming the wait; submitting while a job runs returns "You already have a BFL video generation running."
• The quota is ACCOUNT-WIDE, not per-Hermes-profile. If another profile (e.g. creative) is generating, YOUR submit fails with the "already running" error — and you don't have their job id to poll. Coordinate out-of-band: only one profile submits at a time.
• When another profile owns the generation slot, hand off work as a prompt FILE (both segment prompts + iteration rules in one .txt) rather than competing for the slot. The user relays between profile sessions.
• A job survives client restarts: re-poll the same job id, never resubmit — resubmitting spends duplicate budget.

Behaviors worth knowing (from bfl_flux3_prompting_guide)

• Multi-shot works INSIDE one generation: "SHOT ONE ... HARD CUT. SHOT TWO ..." produces real cuts; "one continuous unbroken shot" gets an uncut take. Consecutive shots must contrast in scale, location, or colour or they blend back into one take.
• Audio lands as separate layers — name ambient, music, and speech separately. Leaving audio unspecified gets invented sound, not silence.
• A quoted line becomes speech ONLY with a visible speaker on camera; otherwise it tends to burn in as on-screen text. Add "no on-screen text, no subtitles" when text is unwanted.

Pitfalls

• Don't chain generations speculatively: each costs minutes and budget. Lock the storyboard first.
• Don't promise a clip length the tools can't make in one pass (>20s needs chaining — say so up front).
• Contests on X: scan what other entrants post so the submission doesn't blend in.
• User taste (suspense/creature trailers): restraint with ONE clear glimpse of the subject (e.g. face held for a beat, then gone) — not pure implication, not full reveal. Confirm the glimpse level before locking a storyboard.