FilmLab

Prompt check of ONE generated image or clip: does the result show what its prompt asked for? The backend first DESCRIBES the result without seeing the prompt, then splits the prompt into checkable claims (counts as digits, quoted text, identity, place, framing, style; for clips the motion) and scores each. verdict: pass | mismatch | unverified, plus `checks`, `issues` and the neutral `beschreibung`. Runs automatically after every generation (lab_job_status shows it as `abgleich`); call this to read it, or with neu=true to check again. On mismatch: offer at most ONE retry per original, lead the new prompt with the failed claim; never re-run a video on your own. Free for the account (FilmLab carries the cost).

Append a shot to a sequence's timeline. Call it once per shot, in cutting order — `start_time` is computed from the clips already there, so appending in order is enough. The continuity pass reads exactly this timeline: a shot missing here is a shot nobody checks. Returns the new clip's index and the sequence's total duration.

Bring your own video, audio or image into a project — not generated, yours. Two ways: `media_url` (a public http(s) address; the server fetches it, max 100 MB) or `file_name` + `content_type` (you get a signed `upload_url` and the `header` that are part of its signature: PUT the file's raw bytes there within the hour with EXACTLY those headers — the answer's `anleitung` has the ready curl line; a missing header is a 403). Either way the answer is the `asset_key` — use it in lab_timeline_apply as clips[].asset_key (video) or bett.asset_key (audio). Exception: in projects that check images, an uploaded image is not in the project after the PUT — the answer then carries `abschluss_noetig` and an `abschluss_key`; after the PUT call lab_add_media again with the same project_id, art, herkunft and `abschluss_key` only. That call checks the image and returns the `asset_key`, or says why it was refused. Free. Internal addresses are refused.

Read a photo and extract location fields from it (Gemini Vision, costs credits — exact price via get_cost: true). Returns the fields — it does NOT create a board; feed the result into lab_create_location yourself. Note the payload: the image travels as base64 inside the call, so keep it small (a downscaled JPEG, not a full-resolution capture).

Animate a completed image into a clip with a video model. Pick the model from lab_list_models; check the price with get_cost or let the user confirm a card via lab_propose. Costs credits when called without get_cost.

Animate MANY images into videos in ONE call (max 25 jobs). Prefer this over looping lab_animate: it validates every job BEFORE spending any quota, so an invalid job 12 of 15 does not leave 11 paid-for generations running. Quota is spent once, collectively (one ledger row); refunds stay per-clip, so a partial failure refunds only that clip. Returns a batch_id — poll lab_batch_status with it instead of polling each id separately. `get_cost: true` preflights the whole batch without generating anything.

Compute the sequence's clip durations from the project's cutting preset and — when the tempo allows it — pull the cut points onto the measured beat grid. WRITES the timeline; costs nothing. The old duration of every changed clip is kept in `clip.metadata.dauer_vorher`, so one undo step exists (a second differing run overwrites it — there is no history). Clip count is never changed: no clip is added or removed. SNAPPING IS NOT GUARANTEED, AND A REFUSAL TO SNAP IS A RESULT, NOT A FAILURE — read `gerastet` and `raster.grund` before you repeat anything: 'tempo_unverlaesslich' = measured below the confidence threshold (deliberate; pass trotz_unsicherem_tempo to override), 'nicht_gemessen' = run lab_beatgrid first, 'keine_musikspur'/'musikspur_nicht_aufgeloest' = run lab_bind_audio first, 'rhythmus_ohne_raster' = the preset's rhythm is 'frei' or 'atmend', which means not snapping IS the promise ('auf den Beat', 'auf die Songstruktur' and 'gegen den Takt' are the three that snap), 'toleranz_null' = you passed toleranz_s 0. In every one of those cases the durations are still recomputed from the preset and written — that alone is the difference between 40 and 6 cuts over 30 seconds. Also read `raster.getroffen` of `raster.von`: `gerastet: true` only says snapping ran; this pair says whether it landed. And `geklemmt` lists clips the renderer's own duration clamp moved back OFF the grid.

Break one scene down into a STRUCTURED shot list of 2, 4 or 6 shots — each with shot size, camera angle, purpose and its own ready prompt, plus one shared floor plan (anchor, axis, light) so the shots cut together. Two families: an 'aufloesung' shows ONE moment from several angles (coverage, dialog, orbit, pov, zwischenschnitte for cutaways …); a 'montage' lets time pass between shots (ursache_wirkung, zeitverlauf, kleiner_bogen …). An unknown art is rejected with the list of known ones. Costs no credits: it is one text-model call and renders nothing; generate the shots afterwards with lab_generate / lab_animate, one prompt each. This is NOT lab_shots_ableiten — that one groups clips that already exist on a timeline; this one plans shots that do not exist yet. By default nothing is written. uebernehmen=true appends the shots as scenes to the project's shot list (mod_shotlist), which the image and video steps read — ask before you set it.

Auto run with a hard spending cap: the person approves ONE run with an upper limit in credits; inside it you may generate without asking per item, but never above the cap. aktion: anlegen (project_id + deckel_credits → status 'angelegt', nothing can be spent yet) · freigeben (ONLY after the person explicitly approved the cap) · stand (cap, reserved, free, items) · beenden. While a run is active, EVERY paid call in that project reserves its price against the cap — whichever tool makes it; reservations are never refunded (a failed item keeps its share, a retry reserves again), upscales and exports count too. The cap counts credits: runs on the person's OWN provider key (BYOK, 0 credits) are not counted. Other projects are not affected. When the cap would be crossed the backend answers budget_exceeded: stop, report what finished, what is open and the minimum budget still needed. Under budget pressure lower the resolution first, never cut runtime, never drop character or location variants.

Auto-generate a storyboard from either script OR situationen (mutually exclusive). An LLM splits script into scenes; situationen bypasses that LLM entirely and generates exactly one frame per entry, ignoring max_scenes. situationen requires figur_id of an existing character sheet with a reference image: every frame uses that character on seedream_5_pro, without creating a new sheet. format selects 16:9 (default) or 9:16 for every model. bewegung_en is saved in each frame’s metadata for later animation. COSTS CREDITS — every scene is one generated image, charged up front for the number of scenes the script is split into (at most max_scenes). Ask the human before running it. CHARACTERS ARE NOT GUESSED: every character who appears in more than one frame first gets a reference image (an existing project character with a reference is reused for free; a new one is created and its reference sheet is charged separately as character_sheet_generate), and every frame showing such a character is generated WITH those references on seedream_5_pro instead of `model`. Frames without a recurring character use `model`. The answer lists `figuren` (name, status, scenes), `figuren_kosten_geschaetzt` and `kosten_je_bild`; prices per model come from lab_list_models. If a reference cannot be made, the frames that need it fail and are refunded — they are never generated without it. The scene split is SYNCHRONOUS and can take up to ~55s and then time out (504) for long scripts; run it in the app if so. The answer lists the new image ids; the frames themselves render in the background as pending images — poll them with lab_job_status (type 'image').

Aggregated progress of a video batch started by lab_animate_batch. Returns {total, completed, failed, pending, done, items[]} — poll this ONCE per batch instead of calling lab_job_status per clip. `done: true` means nothing is pending any more (some items may still have failed). Each item carries retry_count (an automatic second attempt happened) and fallback_from (the provider was swapped after a content-policy rejection).

Turn the still-image clips of a sequence into video clips that exactly fill their beat-snapped slots. Free — this tool itself generates nothing. aktion plan: one job per image clip, ready for lab_animate_batch — start image, the frame's motion (bewegung_en), and the SMALLEST duration the model can deliver that still covers the clip's timeline duration (a shorter video would end early and shift every later cut off the beat). Refuses with 409 if a start image is not accepted yet. Pass the returned `jobs` unchanged to lab_animate_batch (get_cost first, then run). aktion einsetzen: once the videos are done, each finished video replaces its image clip — same duration, source in-point 0. Repeatable; reports what is still open and any video shorter than its slot (left as still). Run lab_beatgrid and lab_apply_rhythm on the sequence BEFORE plan, so the durations are the beat durations.

Measure the TEMPO of the sequence's bound music track and store the beat grid — opens the real file with ffmpeg and analyses it, no guessing and no provider, so it costs nothing but a few seconds. Writes `bpm` and `metadata.beat_grid` on the audio row; it does NOT touch the timeline (`timeline_geaendert: false`). Deterministic and cached: a second call returns the same numbers without downloading again (pass `neu_messen` to force). READ `confidence` AND `verlaesslich` BEFORE YOU CUT. Below the threshold (0.35) lab_apply_rhythm computes durations but deliberately does NOT snap them to the grid — that is a decision, not a failure. Soft material measures badly: a real swing track came out at 0.164, an ordered hard pulse (techno, 128 bpm) at 0.621. Half and double tempo are the classic misreads, so `bpm_alternativen` is part of the answer. `frame_konflikt` reports the collision between beat grid and frame grid, for BOTH renderers: `export` (24 fps, the delivered MP4) and `murnau` (25 fps, the editor timeline). Read `frames_pro_schlag` as a DEVIATION, not a count: it is how far one beat sits from a whole frame, between 0 and 0.5. 0 means beat and frame coincide (100, 125 and 150 bpm at 25 fps do); anything above means beat-accurate and frame-accurate cannot both hold, and the rounding error accumulates across clips — measured on a finished MP4 at 128 bpm: up to 64.7 ms off the audible beat although every cut point had snapped. Shown here, not fixed: that decision belongs to a human. Refusals that name their own remedy: 409 without a bound music track (run lab_bind_audio), 409 when the track is pending (poll lab_list_audio), 502/504 when the audio host is unreachable or slow, 422 when the file carries no usable audio, 503 when ffmpeg is missing.

See a clip at EXACT seconds: up to 12 frames, in the order you list them, tiled into one image and shown. Use it to check what is on screen where you plan a cut, whether a motion has ended, or whether a face drifted between two moments — lab_contact_sheet spreads frames evenly and cannot answer 'what is at 3.2 s'. A second past the clip's end is rejected by name, not clamped. Works on generated clips (video_id) AND on your own uploaded footage (project_id + asset_key from lab_add_media) — look at it before you cut it. Free (no credits).

Bind the project's generated audio to a sequence's tracks — ONE call that does both halves of the job, because doing only the first half silently produces a silent film. Half one (`POST …/soundmix`) creates the four empty tracks (dialog, music, sfx, atmo); half two (`POST …/soundmix/resolve-stems`) is what actually attaches the audio rows to them. On 2026-08-28 that split cost a complete silent export: the mix looked configured, and every track was empty. From here it is one tool, and it is idempotent — an existing mix is NOT re-initialised (that would throw away volumes, effects and the other tracks' bindings), only re-resolved. Per track it takes the audio row marked `selected`, otherwise the newest one with a url. Rows still pending have no url and stay unbound. That choice is re-made on every call: generate a second music track and run this again, and the music track follows the NEWER row — unless one row is pinned with `selected`, which is set in the web UI. With `audio_ids` the named rows are bound and PINNED, and every track you do NOT name is pinned EMPTY (so an old narrator or sfx row of the project cannot slip into the mix); later automatic resolves, e.g. before a render, keep all of that. A call WITHOUT `audio_ids` releases every pin and resolves as described above. Changes state, costs nothing. Next: lab_beatgrid.

See the CUT at exact seconds of the finished film: up to 12 frames, tiled into one image and shown. Each second is mapped like the export maps it — which clip is on screen, where in its source (src_in_s counts), the export's framing (scaled and letterboxed), title cards drawn. A video whose source is shorter than its clip duration ends early in the export, and everything after it moves forward; this tool follows that, so you see the film that will be rendered. The answer names per frame the clip (position, clip_id), clip start and offset; a clip that would render black says so ('schwarz'). NOT in the image: lens/grade, lower thirds, the AI badge — that needs lab_start_export. Use it after lab_timeline_apply to check cut points before exporting. Free (no credits).

Create character sheets for the people found in a project's material — the writing counterpart to lab_cast_gaps. Creates at most 6 per call and binds each to the project. Costs an LLM run per call. Run lab_cast_gaps first if you want to see what it would create. The sheets it creates have NO reference image: each one still needs lab_generate_character_reference before lab_generate accepts it as `character_refs`. Fail-open: no script or no AI key returns an EMPTY `created` list with a `note`, not an error.

Find people who appear in a project's material (script / cast module) but have NO character sheet yet. Returns up to 6 `missing` entries, each with a name and a short hint from the text. It creates nothing — lab_cast_autopopulate does that. Costs an LLM run, so it is credit-gated. Fail-open by design: no script, no AI key, or an unparseable answer all come back as an EMPTY list (with a `note` saying which) — an empty list is the correct answer, not an error.

Build a contact sheet for a generated clip — N frames spread evenly over its runtime, tiled into ONE image — and show it. This is how a clip gets judged at all: MCP has no video content type, so without a sheet a clip is just a URL and a promise. It is also the fastest way to SEE camera drift: measured 2026-08-04, the wander in one clip was a PSNR number until frames sat side by side. The sheet is cached on the video row — a second call returns the same URL without re-rendering (pass force to redo it). Rendering downloads the clip and runs ffmpeg, so the first call takes a few seconds.

Start the continuity pass over a sequence — every shot is compared against its PREDECESSOR and, with check_references, against the character and location reference boards. This is the pass that catches drift (the flat that changed between two shots, the cardigan that turned into a coat), which is the actual enemy in AI film, not the cut. Runs in the BACKGROUND: returns a `job_id` with HTTP 202 — poll it with lab_get_job until status is done|failed, the findings sit in the job result. Two refusals worth knowing before you spend time: an empty timeline is rejected with 400 (add shots with lab_add_clip first), and the backend answers 503 if GEMINI_API_KEY is not configured.

Create a character sheet for a human, animal or creature in the cast register. The backend derives an English `prompt_prefix_en` from the fields you give — that prefix is what later describes this character in every image, so the richer the description, the more stable the identity. This is step 1 of 2: a fresh sheet has NO reference image and is therefore not yet usable as `character_refs`. Run lab_generate_character_reference on the returned id to finish it. Free — it writes a row, it burns no credits.

Create a location board. The backend derives an English `prompt_prefix_en` from the fields you give — that prefix is what later anchors every image and clip shot there, so the richer the description, the more consistent the place. Pass project_id to bind the board to a project (recommended: unbound boards are invisible to lab_list_project_locations).

Create a new FilmLab project. Returns the new project id. Optionally set a title.

Create a sequence (Stage 9 timeline) inside a project — START HERE to turn generated shots into something the continuity pass can read. The new sequence is EMPTY: add shots with lab_add_clip before anything else works. lab_continuity_check rejects an empty timeline with 400. Returns the sequence record including its `id`.

DELETE a location board permanently. There is no undo and no soft-archive — the row is removed. Images and clips that were generated with this `location_id` keep their already-rendered prompt, but the board they point at is gone: nothing can be re-anchored to it, and lab_get_location will 404. Confirm with the user before calling this. Use lab_restyle_location if you only want the place to look different.

Create or read a depth map for a project video in FilmLab: use it for fog, depth of field, relighting or occlusion. Near is white, like Resolve's Depth Map. Free on our own CPU, no credits; processing takes minutes. modell: v2 (default) is Depth Anything V2, sharpest edges. vda is Video Depth Anything: steadier over time (no brightness pumping across the shot, slightly less flicker), softer edges, about 2x slower than v2 (on our server ~0.5 s per frame for v2, ~1.1 s for vda), only for 16:9 landscape or 9:16 portrait videos (16:10 and 1.85:1 are fine, 4:3 and 2.39:1 are not). Each modell keeps its own map; always pass the same modell when reading again (engine in the answer says which one). Reuses the existing map or running job by default; neu: true starts a new job. Returns status and job_id; when done, tiefe_url is FFV1 gray16 and vorschau_url is a playable H.264 grayscale preview, both signed for download without a browser session. wait_s waits at most 15 seconds (default 0). If still running, call this tool again with the same project_id, video_id and modell, neu: false. Once done, use resolve_ergebnis_an_timeline with vorschau_url to put the preview as a matte above the original clip in Resolve.

Is everything wired? Backend reachability and latency, identity, credit check, catalog age, last error of this account, host widget support. Traffic light plus one sentence per point. Read-only, free.

Build a dossier of a project's takes: one ZIP with an offline index.html (no external resources) plus the media files, named Project_SceneNN_YYYYMMDD-Tnn.ext. modus 'kunde' is for a client and NEVER shows seed or costs, whatever the switches say; 'intern' shows everything that helps to rebuild a take. Both modes carry the AI disclosure (EU AI Act Art. 50) on the page and in the image files. Free (no credits). Runs as a job — poll lab_get_dossier.

Dub a sequence into another language: either a ready audio track (en_audio_url — e.g. the one lab_dub_tts just produced) OR a cloned voice reading a line of text (voice_id + dub_text), then a controlled visual lipsync runs over the sequence's own footage (ROI-crop + composite for a long static talking-head shot, or the full-frame veed_v2 path for multi-shot sequences and shots where the mouth is partly hidden). ASYNCHRONOUS: returns a job_id immediately: poll lab_job_status until it completes. Needs a cloned voice first — lab_voice_clone makes one, lab_list_voices shows what a project already has.

Generate the EN-dub audio track for a sequence via ElevenLabs TTS — SYNCHRONOUS, returns a ready audio_url immediately, no job_id and nothing to poll. Optionally pads the track with silence up to the sequence's own length so a shorter reading doesn't shorten the final cut. Feed the resulting audio_url into lab_dub's en_audio_url to lipsync the sequence to this track.

List the ElevenLabs TTS voices available for lab_dub_tts's `voice_id` — this is the EN-dub-track text-to-speech voice list, NOT a project's cloned voices (those are lab_list_voices). Reads only — no state changes, no credits. If ELEVENLABS_API_KEY isn't configured on the backend the list comes back empty and `configured` is false.

Append context to a board's `prompt_prefix_en` (plain string concatenation server-side — no AI call, no quota). Idempotent: context already present in the prefix is not duplicated. This does NOT rewrite or regenerate the prefix — for a board whose prefix is empty, too thin or a raw field dump, use lab_restyle_location (it rewrites prompt_prefix_en with an AI call, costs credits — exact price via get_cost: true) instead of recreating the board.

List the available export delivery profiles (web, social_vertical, social_youtube, streaming_hdr, cinema, master, alpha_prores4444, alpha_png_sequence) and quality presets (draft, standard, high). Call this before lab_start_export to pick a valid profile/quality. The two alpha profiles keep transparency (ProRes 4444, or one RGBA PNG per frame delivered as a zip) and are REFUSED unless the sequence actually carries an alpha source — a cut-out layer element, or a Switch X transform whose alpha track was kept.

Use when the person asks about earlier states of a cut, wants to go back to a previous version, compare versions, name a version, or create a second cut (e.g. a short version next to the long one). The history and the versions of ONE cut (sequence). Every timeline write (lab_timeline_apply, the editor, the Resolve bridge) first saves the state it replaces; this tool reads those saved states and works with them. Costs nothing. aktion 'liste' (default): the saved states, newest first, with `marke` (a name a person gave a state), plus where this sequence was branched from (`ursprung`, with what changed since the branch point) and which versions were branched from it (`abgezweigt`). 'markieren': give a state a name (`marke`); a named state is never cleaned up — states without a name are capped per sequence. Empty `marke` removes the name. 'vergleichen': what changed since a state — against the current cut, or against another state (`gegen`), clip by clip. 'zurueckholen': write a state back as the current cut; the cut it replaces is saved first, so this is undoable. 'abzweigen': make a NEW sequence (a version — e.g. a broadcast cut next to the festival cut) from a state, or from the current cut without `checkpoint_id`. The new sequence keeps the clip ids and the shot layer, knows its origin, and the branch point gets a name so it stays. There is no merging back: a version is its own cut from then on. Edit it with lab_timeline_apply on the returned sequence id.

Tell the FilmLab team about a bug or a missing feature — it reaches the people who can fix it, the chat does not. Use art 'fehler' when a tool returned something wrong, failed without a clear reason, or contradicted its own description; 'wunsch' when a step you needed does not exist. Write what you called, what you expected and what came back; put ids, arguments and the error text into `kontext` (no secrets, no personal data of third parties). One report per problem. Do not report a refused request that was correct (e.g. 'Project not found' for someone else's project). Needs an account. Free.

Finish an accepted DRAFT (a lab_reference_motion run with draft: true) in 1080p — same motion, same seed, same sound, sharper picture. A NEW video row comes back; the draft stays. Works only while the draft is valid and only through the account that ordered it; otherwise the backend refuses with a named reason (entwurf_abgelaufen, entwurf_anderes_konto, entwurf_fehlt, bereits_fertiggestellt) before anything is charged. Never download and re-upload a draft to get here: that is a new run, not this draft. Costs credits when called without get_cost.

Use when the person wants a saved flow, pipeline or multi-step workflow priced, costed, calculated or checked against the balance BEFORE anything runs. Price a whole Flow (Pipeline) before running it: one quote over every node, checked against the credit balance and the project's cost cap. `workflow` is a FilmOS workflow in schema v1 (vault/workflows/_schema.md): nodes of type `lab.tool` call FilmLab tools by name, `{{node_id.field}}` passes one node's result into the next, `parallel` and `condition` are supported, `foreach` and `workflow.invoke` not yet. Every node whose tool costs money is priced with that tool's own get_cost preflight; a paid tool without a preflight, or a price that cannot be determined, rejects the whole quote by name — nothing is estimated. Nothing is generated and nothing is charged. Returns `angebot_id` (valid 24 h, for exactly this workflow, these inputs and this project), the total, one line per node, the balance and what the cap still allows. A node whose price only exists once its input exists (lab_upscale fed by a clip from another node) is listed with `preis_folgt: true` and is NOT in the total: the run stops before it, names the exact price and continues only after lab_flow_fortsetzen — tell the person so when you name the total.

Use when a flow/pipeline run is waiting for approval or a price confirmation and the person gives the go-ahead to continue. Resume a Flow (Pipeline) run that stopped and waits (status 'wartet' with a reason like 'Wartet auf Freigabe'). A run stops when a clip node needs a start image the person has not approved yet — approve the image first (lab_review_asset, verdict accepted), then resume. A run also stops before a node whose price only exists once its input exists (lab_upscale after a clip — the quote listed it as `preis_folgt`): the reason starts with 'Wartet auf Preisbestätigung' and names the exact credits. Resuming IS the confirmation of that price — ask the person first. The confirmed amount travels as max_credits; if the price rises before the run, nothing is charged and the run stops again with the new price. The waiting nodes are tried again; nothing that already ran is repeated.

Use when the person has seen the quote for a flow/pipeline and says to run, start or execute it. Run a Flow (Pipeline) that was quoted with lab_flow_angebot. Pass the SAME workflow and inputs plus the angebot_id — a changed workflow, an expired quote (24 h) or a quote that was already used is refused; get a new quote then. Only after the person confirmed the price. Returns right away with `lauf_id`; the run continues on the server without anyone watching. Follow it with lab_flow_status. Each node calls its FilmLab tool under your account; every tool's own gates still apply (approved keyframe before a paid clip, cost cap of the project). If the cost cap is hit, the whole run stops and nothing after it starts. The quoted price is also the run's own hard cap (auto_lauf_id): every node reserves against it, failed nodes keep their share, and a run whose retries would cost more than the quote stops with budget_exceeded instead of spending more.

Use when the person asks how far a flow/pipeline run is, whether it finished, why it stopped or hangs, or what a step returned. State of a Flow (Pipeline) run started with lab_flow_starten: the run's status (wartet, laeuft, fertig, fehlgeschlagen, abgebrochen) and, per node, its status, attempts, result and error. A node's result is exactly what its tool returned — ids from there go into the next tools. With `angebot_id` instead of `lauf_id` it finds the run started from that offer — `gefunden: false` means no run was created from it, so nothing was charged for it.

Generate an image on the production pipeline (fal.ai, quota-tracked, persisted to the project's gallery). Waits up to `wait_s` seconds (default 45) for the image and then answers with the finished rows — image_url set, inline previews attached, so the picture is in the answer like lab_job_status with preview. Pass `wait_s: 0` to return right away with rows in status 'pending' (poll lab_job_status with id + project_id then). If the budget runs out, the rows come back as they are and `fertig: false` says so. Default tool is 'flux_pro'; use 'nano_banana_pro' (Gemini 3 Pro Image) for the Papercut-style look. Pass seed_image_url to run img-to-img guided by an existing image URL. `get_cost: true` preflights credits — it returns the price without generating anything. FOR A PERSON WHO MUST LOOK THE SAME across several images, do not re-describe them per prompt: get an id from lab_list_characters (or build one with lab_create_character + lab_generate_character_reference) and pass it as `character_refs`.

Generate the reference sheet for a character — six core views rendered from the sheet's `prompt_prefix_en`: detail (three-quarter close-up, the identity anchor), frontal (frontal close-up), profile, fullbody (front), fullbody_back (back view: hair and the back of the costume) and costume (the outfit on an invisible mannequin, NO person — use it where the costume is meant, never as a face reference). Back view and costume are derived from fullbody, everything else from detail. The poses pose_sitting, pose_walking, pose_action are optional extras for blocking — not part of the default run and not used as references for image or video models. A seventh view, 'detail_smile', is NOT in the default set and must be asked for: the same close-up with the mouth open and the teeth visible. Render it for every character who SPEAKS — otherwise the model invents the teeth and the jaw the first time they laugh, and the smile arrives as somebody else's mouth. It is an extra view for the mouth, never an identity anchor: a smiling reference carries its smile into every shot, funerals included. MEASURED 2026-09-15 (n=10, criterion fixed beforehand): the teeth land 10/10, but the PERSON only holds in about 4 of 10 even with the anchor active — two runs were plainly a different man. Generate it, then LOOK before you rely on it; the failure is obvious to the eye (a stranger's mouth on your character), not silent. THIS IS WHAT MAKES A SHEET USABLE: until it has run, lab_generate rejects the id as `character_refs`. Step 2 of 2 after lab_create_character. COSTS CREDITS — one provider call per view; the default run is six for humans, five for tier/kreatur (no costume). Pass `views` to render fewer, and get the exact price first with get_cost: true (free). Takes about 25 s PER VIEW (the tool requests them one by one, anchor first), so a full run is 2–3 minutes. If a single request times out, the answer says `noch_in_arbeit` instead of failing: that view is still rendering and already paid for — check it with lab_get_character_sheet, do NOT re-run it. The sheet needs a `prompt_prefix_en`; the backend regenerates one on demand but only best-effort, and answers 400 if that fails too. Views whose image could not be mirrored to R2 are deliberately NOT saved and come back in `nicht_gespiegelt` — those views are missing, not broken, and a re-run renders only what you ask for.

Generate time-of-day reference images for a location board (costs credits per variant — check the price with get_cost first). Without `variants` the backend renders all four. The images are what makes a board usable as a visual reference.

Generate a MUSIC track for a project (Stable Audio via fal.ai) — the first step of the rhythm chain, and the only one that costs credits. ASYNCHRONOUS: the call returns immediately with a row whose `audio_url` is null and `metadata.status` is 'pending'. Nothing is playable and nothing can be bound yet. Poll lab_list_audio (audio_type 'music') until that row's status is 'completed' and audio_url is set — usually under a minute, and the backend gives up after five (status turns 'failed', metadata.error carries the reason, and the credits are refunded). PUT THE TEMPO IN THE PROMPT if you intend to cut on the beat. This is measured, not assumed: a soft swing track measured at confidence 0.164 (unusable), an ordered hard pulse ('techno, 128 bpm, four-on-the-floor') at 0.621. The tempo detector is autocorrelation over an onset envelope — it finds a hard pulse and shrugs at drone, ambient and rubato. Then: lab_bind_audio, lab_beatgrid, lab_apply_rhythm.

Get ONE character sheet in full — every descriptive field plus the three that decide what you can do with it. `prompt_prefix_en` is the text block describing this person; it is what the backend prepends wherever the sheet is used. `reference_images` decides whether the sheet is usable at all: lab_generate REJECTS a character_refs entry without one, before any quota is spent — if it is empty, run lab_generate_character_reference first. `states` carries the state sheets against character drift (costume, light, physical condition); lab_get_character_sheet renders them agent-readable.

The character sheet PREPARED FOR PROMPTING: base prefix, the state sheets by name, which reference views exist, and the ready-made sentence to paste into a lab_generate prompt. Same source as lab_get_character, different shape — use this one when you are about to generate, and lab_get_character when you want the raw record. State sheets exist against character drift: one sheet per costume, light or physical-condition change, so a person stays the same person when the situation changes. THIS IS THE ANSWER to 'same person, different light': character_refs alone routes to i2i and drags the reference's lighting along, so a drastically different light fights the reference. A state sheet puts the change into the TEXT, where the prompt can win. Free.

Read the accumulated context/output for one pipeline module of a project (e.g. the IDEE, STORYBOARD or SHOTS stage). Needs project_id and module name.

Poll a dossier job by job_id (from lab_dossier). status is queued|processing|done|failed. When done, result.archiv_url is a signed https link to the ZIP that loads WITHOUT a login and expires after 24 h — call again for a fresh one. result.fehlend lists takes that could not be loaded.

Poll an export render job by job_id (from lab_start_export). status is queued|processing|done|failed; when done, the result (output url) is included. Call repeatedly until done|failed. NB: this is the EXPORT poll — different from lab_get_generation. Where the address sits depends on the profile: a rendered film is in video_url, but profile alpha_png_sequence delivers a ZIP of RGBA PNGs and therefore leaves video_url null and puts the address in archiv_url — a .zip under a key called video_url would read as something playable to every client. download_url carries the address in both cases and is the safe one to hand to a user. All result addresses — video_url, download_url, archiv_url and the provenance record herkunft.pdf_url / herkunft.json_url — are signed https links that load WITHOUT a login and expire after 24 h; call lab_get_export again for fresh ones. view_url is the short session-bound form for the web app only. A finished result also carries zielmarkt_checkliste: per target market of the project (see lab_zielgruppe) what the customer must observe when delivering there — empty without markets. Pass it on to the person with the download; it is not legal advice.

Get the status of a background job by job_id (status is queued|processing|done|failed). Poll this after starting a long-running generate/export. Returns the job record; its payload and result are objects (for a clip chain: result.zusammenfassung, result.ketten, result.clips). For an Umzug import (lab_umzug): result.bildzeilen (image_id, datei), result.gesperrt (datei, grund), result.ungeprueft. Pass wait_s (up to 80) to wait inside the call until the job is done or failed instead of polling in quick succession.