FILMLAB · Manual
Step by step: from the empty project to the exported film — in the browser, with IRO, through AI agents and from within Resolve, Blender, After Effects, Final Cut or Avid. Every button is named here as it is named in the app, and before every step that costs money, it says what.
Three things every step after this one takes for granted: an account, a balance — and the habit of reading the cost line before you click.
Under every button that costs money there is a line before you click:
The first number is the price in Credits, the figure in parentheses the euro range (what a Credit costs depends on how you buy — 2.5 to 4.0 cents), the last number your remainder afterwards. If it says up to N Credits, the price is a ceiling (for video, for example, where the delivered length counts).
The cost line is the binding figure. Buttons with a cost line below them no longer carry a price themselves — one figure is enough.
Credits back if nothing is produced. If a paid run aborts at the provider — content filter, time limit, outage — the Credits are returned to your balance. What you see then is an error message with a reason, not a quieter result.
Anyone with an account of their own at fal, Hedra, Replicate or OpenRouter can generate through it. That needs the plan Pro BYOK or Studio.
No silent fallback. If your key is invalid, the provider account empty or throttled, you get an error — not a run on our bill. The key is only checked on the first run, not when saving.
Where it does not apply: synchronization (dub/lipsync) always runs through Credits, because its price is blended from several providers; models reachable only through kie.ai likewise.
Studio includes three seats, you included. Invited people work in your projects with their own account — one access per person, no shared passwords.
If all three seats are taken, it says All seats taken. The window shows additional seats (€29 a month, without Credits), but they cannot be booked yet — it says so. What a member triggers is deducted from the Studio balance and carries their name; returned Credits go back to the balance they were charged to.
Idea or screenplay in → create characters and locations → per scene prompt, image, clip → arrange in the Sequencer → fine-cut in Murnau → export, with a provenance record.
There are two modes for this, switchable at the top of the bar:
Idea in, animatic out. You give a story in a few sentences or a whole screenplay, and the pipeline runs through to a rough animatic. Default on first opening. Good for seeing in twenty minutes whether a story holds.
Step by step. The ten steps of the left bar one by one, each with its tools. This is the way for everything that is meant to become pitch-ready — and the way this manual describes.
Ten numbered steps, below them a Tools block that work across all steps:
| Step | What happens there | Costs Credits? |
|---|---|---|
| 1 Idea | Idea or logline → screenplay draft; Express entry; screenplay import | no (draft yes, low) |
| 2 Screenplay | Scenes and dialogue, analysis, auto storyboard | Auto storyboard yes |
| 3 Location | Locations from the script, boards, moodboards, 360° | Boards and moodboards yes |
| 4 Cast | Characters, character sheets, imprint (trained identity) | Sheets and imprint yes |
| 5 Prompt | Prompt Studio: scene, cast and location become the image instruction | Prompt no, preview minimal |
| 6 Image | Image + VFX: generate variants, select, apply a look | yes |
| 7 Keyframe | Mosaic and tiles, refine, animate a keyframe | yes |
| 8 Video | Video + Sound: clips from images, reference motion, sound | yes |
| 9 Sequencer | Build the sequence, bind sound, snap to the beat, opening titles, render | Render no, opening titles yes |
| 10 Export | Choose a profile, approve, export, publish | no |
| Tools | ||
| Story axis | Scene sequence as an axis: sketch, change location, illustrate scene | Sketch yes |
| Layers | Layered Shot Builder — cut out, extras, composite, parallax | Cards and clips yes, render no |
| Pose Director | Posture and blocking of the figure before the prompt | no |
| Crew | Choose a preset (film or trend), override departments individually | no |
| Capture | Own material and photo to 3D | 3D yes |
| Dub/Lipsync | Foreign-language track, clone voice, dubbing | yes |
| Murnau | Handover to the edit bay | no |
If you know the screening: there the ten stages are called Idea · Screenplay · Location · Crew · Cast · Prompt · Image+VFX · Video+Sound · Sequencer · Export. The bar sorts this differently — Crew is a tool, Keyframe a step of its own. The same way is meant.
Every step follows the same pattern: what you need, what you do, what you have afterwards, what it costs. The labels appear here as they appear in the app.
The images in this chapter come from an example we built through all the steps: “The Last Train” — a platform at night, Lena waits, Jonas gets off. Four shots, two characters, one location, 20 seconds. The same film was made with three video models; the comparison is in step 8.
A story in a few sentences — or a screenplay as a file or text. Credits for the images.
A project with a screenplay, scenes, one image per scene, and a rough animatic in the Sequencer.
An idea, a logline — or nothing at all except a style that interests you.
A screenplay draft with scenes, which sits in step 2.
A screenplay in the project — from step 1, from the import or from Script Studio.
Scenes with characters and locations as anchors for all further steps; on request a finished storyboard.
The locations from the screenplay (step 2) — or your own photos of a real location.
One board per location with images, look and times of day, which serves as an anchor in prompt, image and video.
The characters from the screenplay. For an imprint: photos, and for real people a consent ID.
One character sheet per character (several views of the same person) — the reference every shot is measured against.
An imprint from photos of a real person only with a consent ID. The field Consent ID (LK-…) is a gate, not a checkbox: without an ID from the Likeness register nothing is uploaded (Without a consent ID (LK-…), nothing is uploaded.). If the consent is withdrawn, the path closes immediately.
The imprint costs immediately — “the run is paid even if nobody checks”, it says on the button. First sheet, then test image, then imprint.
A scene, plus cast and location from steps 3 and 4.
An image instruction that carries light, optics, camera and the references of your characters and locations — for step 6.
Blocking instead of description. Where a figure stands in the frame is not described, it is placed: Open Blocking Board (in step 8 and in the Pose Director tool) puts cast and props into the frame as boxes; the description with coordinates falls out as a by-product. Measured, this holds the arrangement markedly better than any sentence — and for camera moves the full move takes six generations instead of 360 (Findings).
A prompt from step 5 — or the shots from “Break down into shots”.
Selected stills per shot, with a look on request, as start images for keyframe and video.
Selected images from step 6 — or an image of your own by URL.
Refined keyframes, and on request the first clip from them right away.
One selected start image per shot. For reference motion: a clip as the template.
Clips per shot, and the audio tracks that go with them — music, sound effects, ambience, dialogue.
The model is the most expensive decision in the whole pipeline. For the same ten-second shot there is a factor of six in price between two models, and the calmest is not the most expensive. That is why there is more text here than elsewhere.
Seedance 2.5 accepts no faces — not as a start image, not as a reference image, not in the reference clip, not even for invented characters. The check is done by the provider, not by us: the run is booked, rejected and the Credits are returned to your balance. For shots with a character: build a still in step 6 and animate it with an image-to-video model that accepts faces (Seedance 2.0, FLUX 3 Video, Kling). For rooms, props, atmosphere Seedance 2.5 is the steadiest we have measured.
Veo 3.1 fast delivers eight seconds when ten are ordered — measured twice, the full length is billed.
Original sound stays only if you say so. Models that generate sound along with the picture replace the speaker's voice with a synthetic one — even if the prompt forbids it. With reference motion using a clip and without “Generate sound”, FilmLab puts the template's sound back automatically (ton: quelle); measured drift 0.044 s. For a close-up while speaking that is enough, for lips in a tight close-up not always — listen closely.
A camera move can be supplied as a path: from Kadrage, from the Blender add-on or through the MCP server (kamerapfad). No video model receives the poses themselves. FilmLab translates the path into the form the model understands: Hailuo gets director commands in the prompt ([Pan left] [Push in]), all the others get a sentence (“Slow dolly in toward the subject…”). After every run FilmLab measures how faithfully the movement in the clip follows the path (path fidelity, 0 to 1).
Calibration run of October 4, 2026: one start image (empty interior), four paths, one clip per model.
| Path | Hailuo 2.3 fast Director commands | Veo 3.1 Lite Sentence | Seedance 2.0 Mini Sentence |
|---|---|---|---|
| Move in | 1,0 | 1,0 | 0,79 |
| Lateral move right | 1,0 | 1,0 | 0,59 |
| Crane up | 1,0 | 1,0 | 0.0 — tilts instead of lifting |
| Pan left, then move in | 0,68 | 0,59 | 0,66 |
It follows that:
One movement per shot is implemented reliably by Hailuo and Veo. Seedance 2.0 Mini confuses crane and tilt and does not hold the direction reliably.
No model carries out two movements in a row. Hailuo only moves in, Veo mostly too, Seedance pans almost the whole time. For multi-part moves split the shot or use reference motion.
Reference motion is the precise way. Seedance 2.0 with a template clip rendered in Blender drove the lateral move at 1.0 and deviated in the image by 0.27% of the width from the poses sent.
Director commands versus sentence: between Hailuo (commands) and Veo (sentence) no difference was visible. A comparison within the same model is still outstanding.
One clip per cell is an indication, not statistics. The numbers apply to the cheaper tiers (Lite, fast, Mini) and one start image. Where a pillar runs through the foreground of the image or the camera looks at a flat wall, path fidelity measures less reliably: in front of a surface a tracking shot looks like a pan. Start images with depth in front of the camera measure most reliably.
On the right in the step is the panel Sound generation with four tabs:
| Tab | Example instruction (placeholder in the field) | Tool | Credits |
|---|---|---|---|
| Music | “Cinematic piano music, melancholic, slow, A minor…” | Stable Audio 2.5, up to 190 s | 17 per run |
| SFX | “Rain on a window pane, quiet footsteps on a wooden floor…” | ElevenLabs SFX v2, 0.5–22 s | 1 per 3 s |
| Ambience | “Quiet Berlin old-building stairwell, a clock ticking…” | Stable Audio 2.5 | 17 per run |
| Dialogue | “Female voice, 30 years old, gentle: 'I am waiting here.'” | ElevenLabs TTS Multilingual v2 | 9 per 1,000 characters |
Created several tracks? Pin the one you want with Select, otherwise the newest wins in the Sequencer. Anyone who later wants to cut the music to the beat writes the tempo into the instruction (“120 bpm, steady four-four time”) — that is the most effective lever on the whole route and costs nothing.
Video: by length, model, resolution — ceiling in the cost line · sound: see tableSelected clips from step 8, plus the audio tracks.
A sequence with order, durations, sound and opening titles — rendered as an animatic or handed over to Murnau.
In the Sequencer, in the VFX bar, the tile Pixelate faces. It works on the rendered sequence: render first, then pixelate. Through MCP the tool is called lab_vfx_preset; with dry_run it shows the price without starting.
The effect is selectable: Mosaic (default), Blur or black area. Resolution, frame rate and sound stay; ProRes stays ProRes. The tool treats masked faces, for example under a balaclava, like all others: it finds the visible part and pixelates it.
Exceptions. Individual people can be excluded, today through MCP, with a preview image per person. After that FilmLab recalculates, free of charge.
Checking. The run comes with a contact sheet of all places where frames were interpolated, the second detector added something, or a longer gap remained open. Look there first.
Pixelated automatically, not guaranteed to catch everything — check before publishing. No detector finds every face in every frame; whether the clip is clean is in the end decided by your eye, not the tool.
A rendered sequence. For an export with generated people or voices: a human who approves.
A film file in the target profile, next to it the provenance record — and on request publication on FrameZero.
| Profile | Color space | Codec | Resolution | Loudness |
|---|---|---|---|---|
| Web / FrameZero | sRGB · SDR | H.264 / MP4 | 1080p | −14 LUFS |
| Social (Instagram/TikTok) | sRGB · SDR | H.264 / MP4 | 1080×1920 | −12 LUFS |
| YouTube | Rec.709 · SDR | H.264 / MP4 | 2160p | −14 LUFS |
| Streaming HDR (Netflix/Apple TV+) | Rec.2020 · HDR10 | H.265 / MP4 | 2160p | −14 LUFS |
| Cinema projection | DCI-P3 · HDR | ProRes / MOV | 4K DCI | −24 LUFS |
| Archive / Master | Rec.2020 · HDR | ProRes 4444 / MOV | 4K | −24 LUFS |
| Alpha master (ProRes 4444, cut out) | Result from “Layers” with transparency, 1080p — for compositing elsewhere | |||
| Alpha sequence (one PNG per frame, as ZIP) | the same as single frames | |||
Ten decibels lie between social and cinema. Anyone who uploads a cinema mix (−24 LUFS) to Instagram gets turned down quietly and sounds thin; anyone who delivers it the other way around fails approval. The same sequence, two files — that is why one export per target.
A broadcaster, a festival, a funding body will ask at some point: what in this film is AI, and who approved it? Every export produces a record as a PDF (to read) and JSON (for machines), with a link to a verification page. It sits next to the film file, not inside it.
What the record does not prove: the timecodes apply to the export from FilmLab, not to the broadcast version — conform, grading and transcode lie in between. It is not a C2PA certificate with a chain to a verification authority, but a record of our own, openly documented. Where the system recorded nothing at generation, it says “not recorded”, no guessing. It is kept in a reduced version for ten years beyond the end of the contract.
The crew is ten departments — direction, camera, light, production design, costume, makeup, grip, edit, colorist, sound — with four to five parameters each. They have a say in every chat and every prompt. A preset sets all of them at once.
A street with twenty people is what a video model spoils most reliably. That is why extras here are not made in the model but as a separate layer over the plate — figures as cards, spread out on the ground, rendered.
Dubbing always runs through Credits, even with your own key — the price is blended from several providers. And dubbed versions made via the short route do not get a provenance record yet.
A take of your own in landscape format, up to 15 s, with your voice and otherwise as little noise as possible. The consent of the person in the image — with yourself that is your own likeness. An image of the creature on your body: the rebuilt start image from the hybrid shoot or a sheet (see below).
The same take in which your head — or your whole body — has become the creature: camera, room, clothing and your voice stay, the face becomes a mask.
The template adopts the prompt templates from Mickmumpitz's “AI Performance Capture 2.2” (October 2026) — head swap, animal on four legs, humanoid creature, head and legs, one person out of several — and sends them to MiniMax H3 through the hybrid shoot. One difference from the original matters: there only the masked area is recalculated, at FilmLab H3 recalculates the whole frame and receives the fence only as text. What that means is in the measurement below.
A take of your own (1280×720, 25 fps, 8.8 s, speaking), excerpt 0–5 s, head swap to a horse, twice: without a sheet and with a sheet as image reference. 50 Credits each.
| without sheet | with sheet | |
|---|---|---|
| Room, camera, clothing | stay | stay |
| Head at the collar | sits like a mask | sits like a mask |
| Expression | only the lower jaw | Lower jaw, at the end a grin from the sheet; eyes, ears, nostrils rigid |
| Hand and glasses at the head | lost | kept, glasses move onto the horse |
| Time base | 25 → 24 fps, 5.0 s becomes 5.17 s; the voice sits underneath, shifted back, and drifts by 0.17 s over 5 s | |
What the technique is at FilmLab — and what it is not. A mask gag that holds: wide shot, short clips, one punchline. Not performance capture for speaking close-ups — the facial expression does not follow the take, because at our end H3 has no second video slot for a face reference. Always supply the sheet: it does not bring the performance, but it tells the model that the take is the source for everything except the head.
If you have a graphics card with 24 GB, you will find Mickmumpitz's free workflow in the ComfyUI catalog (MiniMax H3 Performance Capture 2.2): there only the mask is calculated, camera and facial expression stay real.
IRO reads your project and tells you what comes next, what it costs and which look fits. IRO suggests — only what you click is executed. The conversation costs you nothing.
IRO separates two kinds, and you see the difference on the button:
If a question needs a department, IRO brings it in by name — one of the ten crew roles, at most two per answer. What does not exist, IRO says honestly (“that does not exist”) — tested with “Upscale video to 4K?”.
Right of the edit area are three surfaces — Edit, VFX, Sound — each with an input line, sliders and tool tabs. IRO is in each one, backed by a department: Edit (Tarek Mansouri, film editor), VFX (Anders Marder, cinematography), Sound (Rafael Duarte, composer and sound designer). Operation in chapter 5.
She can take it back — the retraction lands as a correction next to the source, not as a deletion, and carries your name; what you correct, nobody else sees. In the app there is no button for this yet; the way is the MCP tool lab_widerruf (chapter 6).
The in-house edit bay, in the browser: timeline, two monitors, color, VFX, sound, Deliver. The clips from the Composer land in it directly, without a detour through a download folder.
Anyone opening a Composer project must be its owner or a Studio member — until September 28 any account could open any project handed over; that is closed.
| Key | Effect |
|---|---|
| Space | Play / pause |
| J K L | in the Source Monitor: slower / pause / faster · on the timeline: one frame |
| ← → | one frame back / forward |
| I O | Set start / end |
| Del / ⌫ | Delete clip |
| R | Ripple on/off |
| + − · Shift+Z | Zoom · show the whole track |
| Shift+I | Edit Index (the edit list as a table) |
| Ctrl/⌘+Z · Ctrl/⌘+Shift+Z | Undo · Redo |
| Esc | Cancel, close overlay |
The Edit Index (button INDEX or Shift+I) shows the edit list: Track, Src In, Src Out, TL In, Dur, Source, Shot — in timecode, video tracks only.
Saving happens by itself, one and a half seconds after every change. The line in the header says so: not saved → saving… → saved · 14:03:12. Every save writes the edit to where an agent reads through MCP — human and agent see the same state.
When the save line asks a question. If the project holds several sequences and it is unclear which is meant, Murnau does not write back but asks: … — Edit NOT applied: which sequence? ▸. If a save overwrites an agent's work, the line says what is gone (3 clip(s) of the agent overwritten …, checkpoint created) — the state before it sits in the history.
A cue is a single statement from the prompt (“hard cut at 00:01.50, camera on the hero, medium”), tied to a window in the edit. The track answers the question: did what the prompt says happen in the image?
da (it is in the image like that), fehlt (does not occur), anders (something happens, but not that), offen (nobody has looked yet).You cure fehlt with emphasis, anders never. With the first, the instruction was missed — clearer, earlier, repeated. With the second, it was understood and reinterpreted — only rewriting helps there. Anyone who lumps both together tries the wrong remedy until they consider the model useless.
| Surface · Slider | Area · Default | high means | low means |
|---|---|---|---|
| Edit · Pace | 0–200 · 100 | tighten, remove dead time, cut earlier | leave air, let pauses stand |
| Edit · Tension | 0–100 · 50 | cut on the movement, shorten toward the end | exhale, let it run out |
| Edit · Rhythm | 0–100 · 50 | even lengths, on the beat | deliberately uneven, long against short |
| Edit · Flow | 0–100 · 75 | soft transitions, J and L cuts | hard cuts, let the breaks stand |
| VFX · Intensity | 0–100 · 100 | play out fully | hinted |
| VFX · Blend | 0–100 · 100 | blended in — grain, color, light matched | visibly applied |
| VFX · Detail | 0–100 · 50 | hair and edges clean — slower, more expensive | rough |
| VFX · Edge | 0–100 · 75 | hard, precise edge | soft — right for motion blur and hair |
| Sound · Loudness | −24…0 · −14 | louder toward 0 — social | quieter — cinema and archive (LUFS, a standard, not taste) |
| Sound · Dynamics | 0–100 · 50 | keep quiet and loud apart | compress, evenly loud |
| Sound · Clarity | 0–100 · 60 | speech to the front, noise to the back | leave the room unembellished |
| Sound · Depth | 0–100 · 40 | allow reverb and depth | dry and close |
The tool tabs under the sliders (Transitions, Smart Cut, Tracker, Roto, EQ, Compressor, Reverb, Denoise, Ducking …) are partly still grayed out — what is gray does not exist yet. In Murnau there is no price on IRO's buttons; what IRO runs, runs on our machines.
The 70 looks from the Composer (20 original, 50 after film frames) are not these 18 cards — they sit in the Composer's grade companion (step 6) and in the MCP tool lab_grade. If you want a reference look on a clip, apply it to the clip in the Composer (button Look in step 8) and then bring it to Murnau.
| Profile | for what |
|---|---|
| Web (H.264 · 1080p) | MP4 for Vimeo/YouTube |
| Social 9:16 (1080×1920) | Reels/TikTok/Stories |
| Social 16:9 (1920×1080) | IG Feed/FB |
| HDR streaming (HEVC 2160p) | Netflix/Apple |
| Cinema (ProRes 422 HQ) | Master for color/sound |
| Master (ProRes 4444) | Archive/Delivery |
The loudness meter ◉ LOUDNESS (broadcast −23 / web −16 LUFS) shows the mix; the target values of the profiles are in step 10. For alpha exports, HDR metadata and the cinema and archive profiles with a loudness standard, the export in the Composer is the more complete way.
What Murnau does not have, although it is built: a button for OpenTimelineIO. A sequence can be output as OTIO and brought back in (the way back always creates a new sequence) — today only through the interface layer, not through the surface. The exchange in the surface is Load .ffreel / Export .ffreel, its own format.
The same pipeline, operated by an AI agent — Claude at the desk, Claude Code in the terminal. 83 tools, every paid one with the price before the run. Reading is free; generating costs the same Credits as in the app.
Address: https://mcp-lab.thefilmradar.com/mcp. No API key, no copying tokens: the connection runs over OAuth with your FilmRadar account. The agent never sees your password; in the connector settings you can revoke access at any time.
https://mcp-lab.thefilmradar.com/mcp.claude mcp add --transport http filmlab https://mcp-lab.thefilmradar.com/mcp
After that, in the session: /mcp → filmlab → Authenticate.
No account yet? Create one at lab.thefilmradar.com, then connect. You get Credits for generating in the Composer (Pricing).
get_cost: true. The same call with this switch creates no job and charges nothing — it answers with the price in Credits and euros and your balance afterwards (balance_now).Through the bridge facade, the second step is additionally tied to an approval token: the price answer carries a field freigabe that is valid for exactly this call with exactly these inputs, once and for fifteen minutes. An agent that changes the inputs needs a new one.
What an agent may not do: approve an export with generated people or voices — only a signed-in human can do that in the interface. And it cannot upload anything: MCP has no upload tool; your own reference photos go in through the Composer.
Every client delivers the full list with tools/list; it is data, not a constant, and it grows. The groups, so you know where to look (paid = Credits):
| Group | Tools (selection) |
|---|---|
| Projects & pipeline | lab_list_projects · lab_get_project · lab_create_project (optionally with a film preset) · lab_get_context · lab_run_module Credits (script, analysis, storyboard, shot breakdown …) · lab_auto_storyboard Credits · lab_playbook · lab_wissen |
| Characters & locations | lab_create_character · lab_generate_character_reference Credits · lab_get_character_sheet · lab_cast_autopopulate · lab_cast_gaps · lab_create_location · lab_generate_location_variants Credits · lab_restyle_location Credits · lab_location_panorama Credits · lab_location_recon · lab_analyze_location_image |
| Generate | lab_generate Credits (image) · lab_animate Credits (image → clip) · lab_animate_batch + lab_batch_status · lab_reference_motion Credits (clip or references as template) · lab_transform · lab_upscale · lab_list_models (catalog with prices and input roles) |
| Look, depth, layers | lab_grade (70 looks, deterministic, free) · lab_grade_suggest · lab_grade_from_resolve · lab_depth_map · lab_layers · lab_vfx + lab_vfx_status |
| Sound | lab_generate_music · lab_sfx · lab_voice_tts · lab_voice_clone · lab_list_voices · lab_dub · lab_dub_tts · lab_ton_durchlauf (sound effects per shot for a sequence) — all Credits |
| Sequence & edit | lab_create_sequence · lab_add_clip · lab_timeline_apply · lab_bind_audio · lab_mix / lab_mix_render / lab_mix_messen · lab_beatgrid · lab_apply_rhythm · lab_suggest_durations · lab_shots / lab_shots_ableiten / lab_shots_setzen · lab_continuity_check Credits |
| Check & review | lab_review_queue (as a gallery widget) · lab_review_asset (accepted / rejected / selected) · lab_job_status · lab_get_job · lab_jobs_inflight · lab_contact_sheet Credits · lab_diagnose |
| Export | lab_export_profiles · lab_start_export · lab_get_export · lab_list_exports |
| IRO | lab_propose (a suggestion with a price, as in the app) · lab_run_proposal · lab_widerruf (correct a wrong answer) |
Tool responses with images render in Claude as a widget: gallery, videos inline, detail view with the full prompt and parameters; the review buttons on every card write your verdict back to the lab.
lab_create_project with title and preset → lab_run_module script, then analyse.lab_generate with get_cost: true → show the price → again without the switch, with character_refs and location_id, so that figure and location go along as references.lab_animate with the source_image_id of the accepted image; model from lab_list_models (mind the input roles: Seedance 2.5 accepts no faces).lab_job_status at intervals until done; lab_review_asset sets selected.lab_generate_music — write the tempo into the prompt (“118 bpm, steady four-four time”), lab_list_audio until done.lab_bind_audio binds the track to the music track of the sequence.lab_beatgrid measures; watch confidence and verlaesslich — below the threshold it is a suggestion.lab_suggest_durations shows what the snapping would do; lab_apply_rhythm writes. Snaps only with presets whose rhythm snaps. Complete in Findings and in the recipe Cut to the beat.lab_reference_motion without reference_video_url, with tool: seedance_2_5 and referenzen: [{art: ort, url}], plus resolution, aspect_ratio. 480p for viewing (92 Credits per 5 s clip), 720p for delivery (197 Credits per 5 s clip).reference_video_url = the wide shot, use_depth: false (depth pins the geometry); ton stays quelle, the template's sound sits underneath. Only without people in the master.lab_generate, then lab_animate with a model that accepts faces.lab_ton_durchlauf, first with get_cost, then with hinweis (location in one line).The reference path never falls back to another model. If the provider declines (face in the image, content filter), the run ends with the Credits returned to your balance — not in a silent switch to a model you did not choose.
Resolve, After Effects and Blender get their FilmLab connection from one installer. It puts the files in the right place and creates your personal bridge key — no terminal, no Git, no Python.
The key then sits in ~/.filmlab/resolve.env (Windows: %USERPROFILE%\.filmlab\), readable only by you. Resolve from the Mac App Store is detected and supported too — the sandboxed version has its own paths, the installer writes to both.
For Blender, the installer alone is not enough. The add-on ends up in the extensions folder, but has to be enabled in Blender: Edit → Preferences → Add-ons, search for “FilmLab”, tick the box. And: if no Blender 4.2 or later is present, the installer sets up nothing for Blender — the message about it is covered by Installation complete. Anyone who cannot find the tab in Blender should check these two things first.
Too many bridge keys? Too many active bridge keys on the account — revoke one first. means: revoke an old key in the account, then try again.
Updating: the same installer, or in the Resolve bridge the button ⬆ Install update to vX.Y.Z as soon as it appears. Uninstalling: choose Uninstall, optionally Also delete configuration.
Four connections, one principle: the image you are working on goes through the pipeline, the result comes back to where you are. Where the price is shown before the click and where it is not, is stated with each one.
Resolve Studio (the free version does not expose the scripting API), an open project with a timeline, the installer from above.
Generated takes in the Media Pool or directly on the timeline, with recipe and before/after; on request Magic Mask mattes as an extras mask in FilmLab.
~/.filmlab/bruecke-start.log — and for App Store Resolve in the container below it, ~/Library/Containers/com.blackmagic-design.DaVinciResolveAppStore/Data/.filmlab/.resolve_magicmask_tracken, then resolve_magicmask_lesen at intervals until fertig, resolve_magicmask_exportieren, resolve_magicmask_nach_filmlab. From then on the extras render in FilmLab takes the matte as the person mask of this video.Three limits from day one: tracking only runs with the Fusion page open and the tool in the viewer; the “Render completed” dialog halts Resolve until someone clicks it away; a call has a twenty-second limit. Remove logo (Delogo) uses Bria Video Eraser: place a box around the whole logo, mark In/Out. The price is 12 Credits per measured second; only the softly faded mask is inserted into the original.
The bridge runs a local server; the MCP server filmlab_bruecke_mcp.py (found under ~/.filmlab/bridge/ after the installer) talks to it and to FilmLab:
claude mcp add filmlab-bruecke -- python3 ~/.filmlab/bridge/filmlab_bruecke_mcp.py
First call of every session: bruecke_status — names port, host (Resolve or Blender), version, response time. Then resolve_projekt_lesen. Twelve read tools (timeline, clips with index, markers, Media Pool, node graph, render state …) and three that write (track Magic Mask, export matte, resolve_ergebnis_an_timeline) — every writing tool creates a checkpoint first. Individual clips are addressed through the index from resolve_timeline_clips, never by name. A back door for arbitrary code in Resolve does not exist and never will.
Blender 4.2 or later, the add-on from the installer (tick the box yourself), enabled once in Blender.
A Blender camera is a FilmLab shot: images and clips are made from the camera view and sit behind the camera.
The Kadrage app on the iPhone, the Blender add-on with a bridge key and a camera in the scene.
A handheld camera movement: live in Blender, recorded as keyframes. The iPhone sends only position and rotation, no images.
# and reaches no server.What goes over the network: poses run through a small Cloudflare relay (handy.thefilmradar.com) that the add-on queries. The Blender bridge stays on 127.0.0.1. The relay stores no user data and logs no poses. The recording in Blender enters the existing path as quelle: "blender-kamera"; a paid run remains a separate approval with the price beforehand.
Two bridges at the same time are the normal case — Resolve and Blender register separately (bruecke.json, bruecke-blender.json). An agent through the facade takes Resolve first; if the tool list is empty, it is attached to Blender. Anyone who means Blender sets FILMLAB_BRUECKE_PORT to the Blender bridge or closes Resolve.
The MCP page talks about three tabs “Image, Video and Look”. There are two; Look is a chip in the IMAGE tab.
After Effects 22 or later, Python 3 on the system, the installer from above.
A clip from the current comp frame, imported into the project panel.
Two things that are different from Resolve: the price appears here only after the run in the line Cost: …, not before — anyone who wants to know the price beforehand asks in the Composer or through an agent with get_cost. And After Effects freezes during the run, for up to ten minutes; ExtendScript knows no concurrency. Do not kill it.
The Drop from the repository (it is not in the installer), Python 3, two folders.
From every frame in the input folder a finished clip in the output folder — in Avid, on request, automatically into the bin.
python3 integrations/fcp-avid/filmlab_drop.py. The window ◉ FILMLAB DROP names the folders Input (frames from the NLE) = ~/FilmLab/Eingang and Output (finished clips) = ~/FilmLab/Ausgang. Set Tool, Prompt, Duration (s), Start watch.Eingang/verarbeitet, failed ones to Eingang/fehlgeschlagen. Accepted: png, jpg, tif.Every frame in the input folder starts a paid generation — it also says so in the Drop's log. The Drop shows a price nowhere; if you want to know it, ask in the Composer beforehand. There is no panel inside Final Cut or Avid.
The cases that come up most often — and what is behind them. Most are not errors but limits that the app names.
| What you see | Cause | Remedy |
|---|---|---|
| The button is off, below it Costs N — only M left. | Your balance is too low. | Plan window (click on the balance) → buy a pack or change plan. |
| Upgrade plan → on the video model | The model belongs to the full tier (Pro). | Choose a different model or the Pro plan. |
| Run aborted, Credits returned to your balance, reason “content filter” / “likeness” | The provider rejected the image — with Seedance 2.5 every face, including invented ones. | For characters, choose a model that accepts faces (Seedance 2.0, FLUX 3 Video, Kling). Do not reword the description to outsmart the filter. |
| Sheet has fewer views than ordered | Individual views were filtered; the answer names which. | Run the named view again with different wording — geometry instead of label (“weight on the left leg” instead of “dynamic pose”). |
| The same character looks different in the second image | No reference image was supplied, or the feature is not in the description. | Choose the cast in step 5 (the reference comes along); every feature that should persist goes into the description individually. |
| Ten seconds ordered, eight delivered | Veo 3.1 fast shortens, the full length is billed — measured. | A different model; when in doubt Seedance 2.0 or Gemini Omni Flash. |
| The camera moves although the prompt says “stands still” | A matter of the model, not of the prompt. | Change the model — the frame steadiness per model is in the Findings. |
| The speaker's voice in the clip sounds foreign | The video model regenerated the sound and replaced the voice. | Reference motion without “Generate sound” (then the source sound is put back) — or lay the original sound underneath in the edit. |
| Measure tempo / Snap to the beat are off | The music track carries no audio file. | Init mix, then Connect audio tracks; the reason is shown under the buttons. |
| Snapping reports rhythmus_ohne_raster | The preset carries frei or atmend. | That is intended; choose a preset with a snapping rhythm (e.g. mehrbild, kontrolliertes-chaos). |
| Extras: not trackable | The plate cannot be tracked reliably — AI plates are harder for this than real footage. | A still as the plate, or generate a calmer move. |
| Extras too small / floating | No scale set; the calculation assumes eye level. | Mark the main figure, enter the height. |
| Murnau: Edit NOT applied: which sequence? ▸ | Several sequences in the project, target unclear. | Click ▸ and choose the sequence — Murnau no longer guesses. |
| Murnau: Export blocked — EU AI Act Art. 50 | A consent is missing. | 🛡 Manage consent. |
| Murnau: the check page does not appear right away | The record is created after the render. | Wait up to a minute; the button ✓ Check page appears by itself. |
| Murnau does not open a Composer project | You are neither the owner nor a Studio member. | Have the owner invite you (team seats). |
Agent: Call without freigabe first | Paid call without a price query. | That is how it is meant: price first, then with freigabe. |
| Agent: FilmLab time limit reached; may still be running | The run takes longer than the client waits. | lab_job_status / lab_get_job — the job keeps running; do not restart it. |
| Bridge: No running FilmLab bridge found | Panel was never opened. | Resolve: Workspace → Scripts → Utility → FilmLab; Blender: tab FilmLab, Bar. |
| Bridge: tool list is empty | The bridge is attached to Blender, not to Resolve. | Open the Resolve panel or set FILMLAB_BRUECKE_PORT. |
| Bridge: a tool is missing even though it is new | Resolve is running with old bridge code. | Restart Resolve — a second click on the menu entry does not load new code. |
| Bridge: DaVinci Resolve Free does not expose the scripting API | Resolve without Studio. | Studio is a prerequisite. |
| Magic Mask tracks 0 frames | Fusion page not open or tool not in the viewer; or a stale point after “Clear Strokes”. | Open the Fusion page, put the tool in the viewer, click the figure again. |
| Blender: no FilmLab tab | Add-on not enabled, or no Blender 4.2 or later found (the message was covered). | Edit → Preferences → Add-ons → “FilmLab” → tick; otherwise update Blender and run the installer again. |
| Installer: Too many active bridge keys on the account | Maximum number reached. | Revoke an old key in the account. |
| After Effects does not respond for minutes | The run blocks AE — for up to ten minutes. | Wait. Do not quit. |
| Own key: error instead of a run | Key invalid, account empty or throttled — there is no silent fallback to Credits. | Check with the provider; or run the job through Credits on purpose. |
If that does not help: [email protected] — with the project ID (shown in the dialog Projects) and, if it is about a run, the job number from the cost line or from lab_get_job.
Prices in euros, including tax, queried from the live system on 2026-09-28. Generation is billed through Credits — or through your own key at the provider.
| Plan | Price / month | Credits | Includes |
|---|---|---|---|
| Free | 0 € | 50 / month | For trying out: full planning for one project (idea, screenplay, analysis, storyboard, shot list) and 50 Credits a month that do not roll over. |
| Creator | 19,99 € | 500 | Unlimited projects, export without watermark, funding export profiles. |
| Pro BYOK | 29,00 € | 0 | The platform; generation runs through your own keys (chapter 1). Models without your own key through Credit packs. |
| Pro | 49,00 € | 1500 | Plus the full tiers of the video models (Veo 3.1 Standard, Kling 3 Pro, Seedance 2.5, HappyHorse, Wan VACE), Pose Director and Coherence Sync. |
| Studio | 149,00 € | 6000 | Everything from Pro with four times the Credits, own keys without a second subscription, three seats for the team (chapter 1). |
Change plan or buy packs: click the balance at the top right → Choose → on the plan, Manage sub → for the current subscription. The balance also applies in Murnau: one subscription, one account.
| Run | Credits |
|---|---|
| Storyboard sketch | 1 |
| An image, depending on the model | 2–13 on the reference surface |
| Extras card · extras clip (6 s) | 4 · 16 |
| Sound effect (3 s) · speech (1,000 characters) · music/ambience (one run) | 1 · 9 · 17 |
| Video clip | by length, model, resolution — ceiling in the cost line |
| Seedance 2.5 reference path, 480p · 720p | 92 · 197 per 5 s clip |
| Grade, edit, render, export, tracking, masks, depth map | 0 |
What a Credit is in euros depends on how you buy: between 2.5 cents (Studio) and 4.0 cents (Creator or the smallest top-up pack). That is why every cost line names both. Purchased balance does not expire, the one from the subscription does (Terms § 7: Credits are valid for three months).
Priority. Pro, Pro BYOK and Studio get priority in our own queue — render, tracking, masks, depth, animatic, lipsync, dubbing. At the model provider there is no priority; image and video generation waits there in its order.
Why the cost model looks the way it does — the test day for $45 where $7.50 would have been enough, the factor of six between models — is in the Findings.