# Cinematic Lab {% hint style="info" %} **Cinematic Lab makes still images, not video.** This is the photography stage of your pipeline. Once you have a still you love, take it into **Motion Director** to animate it, **Multi-Cam** to generate alternate angles, or **AI Transitions** to bridge two stills into a moving shot. {% endhint %} #### When to Use Cinematic Lab Cinematic Lab is the right tool any time you'd rather **describe** a frame than shoot it — and you want it to look like a frame from a movie, not a stock render. * **Concept frames and mood boards** — pitch a look before you book talent, locations, or gear * **AI key art and thumbnails** — generate cinematic stills sized to whatever aspect ratio your edit needs * **Pre-vis stills for an edit you haven't shot yet** — drop a placeholder that already matches the lens and lighting plan * **Source frames for video generation** — Cinematic Lab → Motion Director, or Cinematic Lab → AI Transitions, lets you start any motion workflow from a still you actually like * **Reference-driven scene matching** — feed in a still from your edit and ask Cinematic Lab to render the next angle, alternate beat, or styled variation * **Title cards, lookdev plates, story frames** — any time the brief is "make this feel like a real frame from a real production" {% hint style="warning" %} **Cinematic Lab is not a chatbot or a Premiere automation.** It runs as a standalone Studio workflow. When you click Done, your selected images are inserted into the chat as a normal image message — from there you can drag to your project, save them to Library, or feed them into another Studio workflow. Cinematic Lab does not edit footage on your timeline. {% endhint %} #### Getting Started **Step 1: Open Cinematic Lab** 1. Open Chat Video Pro (Window → Extensions → Chat Video Pro) 2. Click the **Studio** button in the sidebar to open the Launchpad 3. In the **Production** department, click the **Cinematic Lab** card Cinematic Lab is one of the few Studio workflows that **does not** open the asset loader first — because it generates from a written description rather than transforming an existing clip. You'll go straight into Step 1: Describe Your Scene. ***

**Step 2: Describe Your Scene** This is the most important box on the page. The model treats your description as the **subject and action** layer of the prompt — everything visual (camera, lens, lighting, photorealism anchors) gets layered on top in Step 3, so you don't need to write technical language here. You need to describe **what's happening, who's in frame, and where they are**. **Three things to always include:**

What	Example
Subject	"A woman in a wool coat", "An empty diner booth", "A weathered fisherman"
Action / state	"lighting a cigarette", "mid-laugh", "staring out the window"
Setting and time	"in a fog-soaked harbor at dawn", "in a neon-lit ramen shop after midnight"

**Optional but powerful:** * Mood and emotional tone ("melancholy", "tense", "intimate") * Wardrobe, props, and color palette ("amber Carhartt jacket, oxidized brass rings") * Weather and atmosphere ("rain, blown sideways", "thick fog catching the headlights") * Composition direction ("close-up on hands", "wide shot, subject lower-third") **Example prompt:** > "A detective standing in a rain-soaked alley at night. Trench coat dripping, neon signs reflecting in the puddles, shoulder of a passing pedestrian blurred in the foreground. He's mid-thought, looking at something off-camera." {% hint style="info" %} **Fastest way to prompt:** Use the microphone button to dictate the scene out loud. Cinematic prompts come out better when you describe a moment instead of a list — speaking is naturally narrative. {% endhint %}

**Reference Images (Up to 14)** Can attach up to **14 reference images** from any of three sources: * **Upload** — JPEG, PNG, or WebP from your computer * **Frame capture** — pulls the current frame from your active Premiere Pro timeline (great for matching a look that already exists in your edit) * **Recents** — pick from any image generated, captured, or used recently across Chat Video Pro. Multi-select is enabled so you can grab a whole mood board at once. References change what Cinematic Lab is doing under the hood. With **no references**, the model generates from scratch using your description plus the camera/lens settings. With **references attached**, the model treats them as a style and identity guide — matching wardrobe, faces, lighting palette, art direction, and texture — while still applying your selected camera and lens characteristics. **What to use references for:**

Goal	What to attach
Match an actor's face across shots	One or two clean head-and-shoulders portraits of the subject
Match a wardrobe or prop	A close-up of the costume / item from any angle
Match a location's color and lighting	One or two stills from the location at the right time of day
Match a film's overall look	3–6 frames from a reference film, all from the same chapter or mood
Generate the next angle from your edit	Capture a frame from your timeline and ask for "the same character from a low angle"

**Step 3: Choose Your Look** This is the page that separates Cinematic Lab from a normal text-to-image generator. Instead of dumping technical jargon into your prompt, you make four choices the way a DP would. The system translates each one into specific visual language the model has been trained to interpret physically — not as filters, but as optical consequences. **The four columns** * **Camera Body** — the *sensor and color science*. Different cameras render skin tones, highlights, contrast, and grain in fundamentally different ways. * **Lens** — the *character of the glass*. Modern primes are clean and clinical; vintage lenses bloom and flare; anamorphics give you oval bokeh and 2.39:1 widescreen feel. * **Focal Length** — *how the world is compressed*. 24mm exaggerates space; 50mm is the human eye; 135mm flattens everything for portrait compression. * **Aperture** — *how much is in focus*. f/1.4 is paper-thin; f/8 is everything sharp from foreground to horizon. You're not picking metadata — you're picking the lens you would have shot on. Underneath, the prompt engine assembles a four-pillar prompt that puts technical setup first (so the model treats your gear choice as the *visual container* before it draws anything) and adds photorealism anchors at the end (so it doesn't drift into illustration). *** **Camera Body — pick the look, not the brand**

Camera	Visual Signature	Reach For When…
ARRI Alexa 35	Soft highlight rolloff, organic grain, exceptional skin tone fidelity, slightly desaturated teals	Drama, narrative, prestige TV, anything where skin and emotion need to land. The default Hollywood look.
Sony Venice 2	Ultra-clean low-light, neutral-cool color, sharp subject separation, clinical clarity	Night scenes, urban exteriors, modern commercial, documentary. Best when you want everything legible in the dark.
RED V-Raptor	Clinical sharpness, 8K-feel detail, punchy reds, high contrast, "digital" precision	Sci-fi, action, VFX-heavy, commercials with crisp products. The opposite of soft.
Blackmagic Pocket 6K	High saturation, thick color density, gritty textured grain, raw indie energy	Music videos, indie shorts, anything that should feel a little unpolished and alive.
Canon C500 Mark II	Warm magenta-pink skin bias, soft warm highlights, organic but sharp	Portraits, interviews, intimate drama, any human-centric frame where the face is the subject.

*** **Lens — pick the personality** Lenses are split into three categories. The category alone changes the **shape of light** in the frame. **Spherical (modern, clean)**

Lens	Best For
Zeiss Supreme Prime	Clinical perfection. Modern features, sci-fi, anything where "razor-sharp" is a virtue.
Cooke S4/i	The "Cooke Look" — warm, gentle focus falloff, faces beautifully. Drama, romance, character-driven scenes.
Sigma Art	Modern, affordable cinema clean. A solid neutral lens when the camera body is doing the heavy lifting.
Canon CN-E	Sharp with Canon warmth. Pairs well with the Canon C500 for skin-first work.

**Vintage**

Lens	Best For
Canon K35 Vintage	1970s gold flares, dreamy bloom, low contrast. Period pieces, nostalgic moods, anything that should feel like a 35mm print.

**Anamorphic**

Lens	Best For
Panavision C-Series Anamorphic	Oval bokeh, horizontal flares, vintage warm character. The classic widescreen blockbuster look.
Cooke Anamorphic /i	Modern anamorphic — clean squeeze, controlled blue flares, sharp corners. Premium widescreen without the rough edges.

**Focal Length — pick the spatial feel**

Focal	Effect	Use For
24mm	Wide, expansive, geometric distortion at edges, deep depth of field	Establishing shots, environmental storytelling, wide group scenes, any moment about the place
35mm	Documentary feel, slight compression, natural human eye view	Interview frames, walk-and-talk, conversational two-shots, "fly on the wall"
50mm	Classic normal lens, zero distortion, what the eye sees	Anything that should feel honest and unstyled — your safest default
85mm	Portrait compression, flattering faces, pronounced bokeh, isolation	Portraits, tight character moments, anything where the subject must read first
135mm	Telephoto compression, flattened planes, extreme bokeh, intimate distance	Long-lens character beats, voyeur framing, dramatic isolation across distance

*** **Aperture — pick the depth**

Aperture	Effect	Use For
f/1.4	Paper-thin focus, dreamy creamy bokeh, subject pops off background	Hero close-ups, mood pieces, anything where the world should melt
T1.5	Cinema-grade T-stop equivalent of f/1.4 — slightly more controlled, more "professional"	When you want shallow focus that reads as cinema rather than photography
f/2.0	Beautiful bokeh, balanced sharpness, classic cinematic	The shallow-focus default — most narrative work lives here
f/2.8	Professional balance, slight environmental context, sharp on subject	Commercial work, two-shots where both subjects need to be sharp
f/4.0	Moderate depth, sharper overall, more world visible	Storytelling shots where the location matters as much as the subject
f/8	Deep focus, sharp foreground to background, hyperfocal	Landscapes, architectural frames, anything where the whole scene is the subject

***

**Output Controls — Quantity, Aspect Ratio, Resolution / Quality, and Model** Below the four selectors is your output bar. * **Quantity (1–4)** — generate up to 4 variations in one batch. With Nano Banana 2 each generation is fast enough that you should default to 4 for first passes — pick a winner, regenerate with adjustments, narrow down. * **Aspect Ratio** — 10 options including 16:9 (landscape default), 9:16 (vertical), 1:1, 21:9 (cinematic), 4:3, 3:2, 2:3, 5:4, 4:5, 3:4. Pick the ratio of the **deliverable**, not the camera — the model renders a real frame for that shape. * **Resolution / Quality** — what shows here depends on the model you choose: * **Nano Banana 2** exposes **1K, 2K, 4K** as a resolution dropdown. Default 2K; bump to 4K when you're picking finals or generating key art. * **GPT Image 2** exposes a **quality** dropdown (Low / Medium / High) instead. High = slowest, sharpest, most detailed. * **Model picker** — a small dropdown lets you switch between the available image models. (See **Choosing the Right Model** below.) When everything's set, click **Generate**. You'll move to Step 4: Results. {% hint style="warning" %} **Regenerating wipes existing results.** If you've already generated a batch and you change anything in Step 3, hitting **Generate** again replaces the previous set. If there's a frame you want to keep, hit **Done** on it first (which inserts it to chat) before you regenerate. {% endhint %} #### Choosing the Right Model **Nano Banana Models** Google's latest image model. * **Fast** is the default. Twice as fast as the previous generation, half the price, supports up to 4K resolution, and is good enough for \~90% of work. * **Pro** adds extended reasoning and web search. Use it for harder prompts: rare locations, real-world references, complex multi-subject scenes, or any prompt where the model needs to "think" about what something actually looks like. **Best at:** * Photorealism — texture, grain, skin pores, lighting physics * Reference-guided generation (reference images influence both content and look) * High-resolution output (4K natively) * Speed — first-pass exploration is dramatically faster than alternatives **Choose Nano Banana 2 when:** you're doing most of your work. It's the right default for portraits, landscapes, character work, mood pieces, anything reference-driven, and any frame headed for video generation in another Studio workflow. **GPT Image 2** OpenAI's high-fidelity image model. **Best at:** * **Text in the image** — readable signs, posters, book covers, neon, packaging, type-driven shots. This is the single biggest reason to switch off Nano Banana. * **Complex multi-subject scenes** — three people in a room, a busy street, a crowded restaurant * **Prompt adherence on weird, specific requests** — surreal compositions, hard logical setups, unusual perspectives * **Conceptual / illustration-leaning frames** — when "real photograph" isn't the only goal **Choose GPT Image 2 when:** there's text in the frame, the scene has a lot of moving parts, or the prompt is conceptually weird. It's slower and more expensive than Nano Banana 2 Fast — but it earns it on the prompts it's better at. {% hint style="info" %} **Pro tip — generate the same prompt on both models.** When you're locking in the look for a project, run the same scene through Nano Banana 2 and GPT Image 2 once. Side-by-side, you'll know within thirty seconds which model owns *this* project's aesthetic. Then commit and stop second-guessing. {% endhint %} ***

#### Working With Your Results * **Click an image** to open it fullscreen for inspection. Click the backdrop or close button to dismiss. * **Click the checkbox in the corner** to add or remove an image from your final selection. The first image is pre-selected by default. * **Regenerate** runs the same configuration again to produce a new batch. (See the warning above about wiping previous results — if you want to keep a frame, click **Done** on it before regenerating.) * **Done** inserts your selected images into the chat as a normal image message and closes Cinematic Lab. **What "Done" Actually Does** Selected images are added to the chat as an image grid (the same layout used everywhere else in Chat Video Pro for multi-image generations) and saved to the Library with full metadata — your scene description, camera, lens, focal length, aperture, model, and resolution all stay attached. From the chat message you can: * **Download the image to use** — Imports directly into Premiere Pro bins * **Pull it back into another Studio workflow from Recents** — it'll show up in the Recents picker for Motion Director, Multi-Cam, AI Transitions, Relight Scene, and the next Cinematic Lab session *** #### Pro Tips **Pair the camera with the lens like you're packing a kit, not picking from a list** Don't shop the dropdowns alphabetically. Decide what the project should *feel* like, then pick a body and lens that already work together in the real world. ARRI + Cooke is prestige drama. Sony + Zeiss is modern commercial. RED + Sigma is "shot yesterday." Blackmagic + K35 is grungy and nostalgic. Canon + Canon is the warmest human-centric pairing in the lab. The model has been trained on actual footage from these combinations — leaning into a real-world pairing produces noticeably more cohesive frames than mixing randomly. **Lead with subject, not gear** The camera/lens controls are doing the technical work for you. Don't waste your scene-description box repeating "shot on Alexa, 50mm, f/2.0" — you've already selected that. Use those words for *who* and *what*. The four-pillar prompt engine layers the technical pillar first under the hood, and the model treats it as the visual container before it draws your subject. **Use 4-up batches as a focusing exercise** Default to quantity 4 for a first pass on any new prompt. Don't pick the best one — pick the *direction* the best one is pointing. Then change one thing (a different lens, a different aperture, a slightly tweaked description), generate 4 more, and compare. You'll converge on a final frame in 2–3 batches that would have taken 15 single-shot generations. **Reference images are the fastest way to character-consistency** If your project has a recurring character, location, or product, attach 1–2 clean reference images or an element on every generation in that project. Nano Banana 2 reads references as both *what to render* and *how it should look* — you'll get cross-shot consistency that's nearly impossible from text alone. For projects with a returning subject, drop a portrait reference into a Library collection once and re-attach it from Recents every session. **Match-the-edit workflow: Frame Capture into Cinematic Lab** If you've already shot something and need a missing angle, an alt take, or a stylized variation, click **Frame Capture** in Step 1 to pull the current frame from your Premiere timeline as a reference. Now describe the variation you want ("same character, low angle from below") and the model will use the captured frame as the identity and lighting anchor while honoring your new prompt and gear choices. This is the single fastest path from a real edit to a believable AI-generated companion frame. **Bridge Cinematic Lab into a moving shot** A still you love isn't the end — it's the start. Once you have a hero frame, hit **Done** to drop it into chat, then: * Drag it into **Motion Director** to animate it with a camera move (push in, orbit, dolly back) * Drag it into **Multi-Cam** to generate alternate angles from the same instant * Pair it with another still in **AI Transitions** to interpolate motion between them * Drag it into **Relight Scene** to change the lighting mood while keeping composition Cinematic Lab is the photography stage — it's most powerful when you treat it as the entry point to the rest of the Studio. **Try GPT Image 2 the moment you need text in the frame** Nano Banana 2 will hallucinate signs, packaging, and titles into something that *looks* like text but isn't. The instant your scene includes a readable word — a license plate, a billboard, a book cover, a chyron — switch the model picker to GPT Image 2. It's the only model in the picker that handles real text reliably. **Pro mode is for hard prompts, not all prompts** Nano Banana 2 Fast is the right default. Save Pro for prompts that need *thinking* — rare real-world locations, weird logical setups, scenes where the relationship between elements has to be physically consistent. For everyday hero frames, Fast is genuinely better because the speed lets you iterate more. **Build a project lookbook once** The most consistent-looking AI projects come from the same trick every time: pick your camera + lens + 4–6 reference images at the start of the project, save them somewhere (a Library collection, a folder, even a chat message), and reuse the exact same set on every Cinematic Lab generation for that project. You're not generating images — you're operating a virtual production with consistent gear and art direction. *** #### Troubleshooting **Generations look painted, plastic, or "AI-y"**\ This is illustration drift. Try to drop the resolution to 2K (4K can over-smooth at certain prompts), and add concrete physical details to your scene description — *"stubble, sweat on his temple, dust on the lapel"* gives the model the textures it needs to stay grounded. **Faces look generic or keep changing**\ Add a reference image. Even one clean head-and-shoulders portrait will lock the face. For projects with a recurring character, drop the reference into Recents once and re-attach it on every generation. **The lens choice doesn't seem to be doing anything**\ Two likely causes. First — your scene description is fighting it. If you ask for "everything in focus" while choosing f/1.4, the model has to compromise. Match your description to your selected aperture, or remove focus language and let the gear do the work. Second — your prompt is too short. The lens characteristics layer in *underneath* a real subject; "a guy" plus a Cooke S4/i won't show off the Cooke Look. Give the model something to render before you expect lens character to appear. **Text in the image is gibberish**\ Switch the model picker to GPT Image 2. Nano Banana 2 is fast and beautiful but it will not reliably render real words. **Generations are too saturated, too HDR, too "Marvel"**\ Try a different camera body. Avoid RED if you don't want punchy contrast. ARRI and Canon C500 are your most natural-looking bodies. You can also bump the aperture from f/1.4 to f/2.0 or f/2.8 — extreme shallow focus can read as artificial. **My references are influencing too much (everything looks like the reference, ignoring my prompt)**\ You probably have too many references attached, or they're too visually similar. Drop to 2–3 references and make sure your scene description is detailed enough to give the model something *new* to construct. **My references are influencing too little (the output ignores them)**\ Nano Banana 2 handles references better than GPT Image 2 — switch models. Also check that your description doesn't directly conflict with the reference (asking for "blonde hair" when your reference has black hair forces the model to choose). **I keep losing my favorite frame when I regenerate**\ Always click **Done** on a frame you want to keep *before* hitting **Regenerate**. Done inserts it into chat where it's safe; Regenerate replaces the current results grid. But all generations are stored in your Fal library even if not saved locally **The Generate button is disabled or generation fails immediately**\ Cinematic Lab needs your FAL API key configured in Settings to generate. If it's missing or invalid, generation will fail before it starts. Set it under Settings → API Keys. Also check that it is funded. *** **Next:** Learn about Motion Director — animate any Cinematic Lab still into a moving shot with a chosen camera move. --- # Agent Instructions: Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter: ``` GET https://docs.chatvideopro.com/features/studio/cinematic-lab.md?ask= ``` The question should be specific, self-contained, and written in natural language. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.