DVP · Core Suite · Workflow Guide

The AI
Filmmaker's
Pipeline

One production flow: Prepare → Image → Video → Vault. Start from an idea, script, or reference frame; prepare the identity and prompt, approve the first image, animate that image, then save the reviewed recipe. CastLock records remain local, and Vault does not yet claim durable CastLoop media attachments.

Start with CinePrompt → See the Workflow
Prepare → Image → Video → Vault + Finish
00 Script to Shotsoptional → 01 CastCastLock → 02 CinePrompt → 03 DVP Vault ⇢ 04 AI Upscaler
How the Production Stages Connect
Optional first
Script to Shots
Break down
Start from writing rather than a blank prompt. Paste a scene, script excerpt or story beats and get one shot per row, each with its own prompt. Skip it if you already know the shot you want.
→ Prepares the shot Shot list · one prompt per row
→
Prepare
Cast
Prepare identity
Identity, in one place. Forge the reference panels a video model can take as input, build the character as a versioned record — identity spine, immutable traits, wardrobe — train a Consistency Model from those same panels, and hold the record across every shot in a sequence.
→ Supplies image evidence Versioned record · reference panels · prompt block
→
Prepare
CinePrompt
Prepare prompt
Full cinematic control — camera, lens, aperture, lighting, grade, texture — shaped for the platform you are shooting on. Build a single shot, translate an existing prompt, or generate a locked multi-shot set from the character you cast.
→ Guides image and video Platform-ready prompts · single shot or multi-shot set
→
Vault
DVP Vault
Remember
Store approved prompt recipes, reference IDs, character notes, and seeds. It is the project memory layer, but the current CastLoop handoff does not promise durable image or video attachments.
→ Project Output Recipe and lineage · exportable
Optional first step

Build Your Shot List

Paste a scene, script excerpt, or story beats. Script to Shots breaks it into individual shots and writes a platform-optimized prompt for every one — ready to generate in Kling, Runway, or wherever you work.

00
Tool
Script to Shots
Any format works — screenplay, prose, bullet points, rough notes. Set the pacing, camera style, and target platforms. Get a complete shot list with one prompt per shot per platform.
Script to Shots
"INT. CAFE – NIGHT
Sarah enters. She
scans the room..."
↓ generating shot list
#01 WS — Sarah enters cafe
#02 CU — she scans the room
#03 POV — target spotted
3 shots · prompts ready for Kling, Runway, Veo
What you do
  • Paste your script or scene text
  • Set visual style (or leave blank for auto-detection)
  • Choose pacing — minimal, standard, or detailed
  • Select your target platforms
  • Click Generate Shot List
  • Each shot card shows a complexity badge (Simple / Medium / Complex — AI-rated 1–3) and a camera movement hint
  • Copy individual prompts, download the full list as .txt, or export to CSV with complexity ratings
Tips for best results
From Cast
Cast a character first and its identity spine travels into every shot you break down here, so the shot list describes the same person from row one to row forty.
Pacing guide
Minimal = 1–3 key shots per beat (action, commercial). Standard = 3–5 shots (drama, narrative). Detailed = full coverage of every moment (cinematic, complex scenes).
Open Script to Shots →
Step One

Cast the Character Once

A video model cannot keep a face it has never seen, and a character retyped into forty prompts drifts forty times. Cast does the whole identity job in one workstation: forge the reference panels as real images, turn them into a versioned record, train a Consistency Model from those same panels, and hold the record across every shot in the scene.

01
Tool
CastLock
Four rooms in one workstation. Forge generates the hero image and builds every reference panel from it — turnaround, expressions, clothing macros — as clean single-subject images a video model can take as input. Build is where the character exists as a versioned sheet — spine, immutable traits, wardrobe, those panels, every revision kept. The LoRA lab trains a Consistency Model from the panels with no re-uploading and checks the model actually renders the same person. Sequence is where a shot list inherits the record, so shot 14 describes the same person as shot 1. Everything is stored locally in your browser.
CastLock · Build
Character · v3
Spine: "Maya Okonkwo, 34, cropped natural hair, chipped front tooth"
Immutable Traits
Face ✓
Hair ✓
Build ✓
Wardrobe ✓
Prompt block · inherited by 18 shots
What you do
  • Open Forge, describe the character, prop or location once, and generate the hero image with your own Gemini key — look at it before running the rest, everything is built from it
  • Tick the panels you want, check the cost estimate, run the batch, and download the panel PNGs or the composed bible sheet; the record lands in the library with no import step
  • Open Build: the forged record is already there
  • Write the identity spine in one sentence, then mark the immutable traits that must never drift between shots
  • Add wardrobe and variants as their own entries rather than editing the character over the top of itself
  • Save a version whenever the look changes — old revisions stay, so you can go back to the one that worked
  • Open the LoRA lab to train a Consistency Model: the training set pre-fills from the panels you forged, so no manual uploads; then compare and drift-check to prove the model renders the same person
  • Switch to Sequence, load your shot list, and attach the character — every shot inherits the same prompt block
  • Send the scene on to CinePrompt for platform shaping, or the record to DVP Vault as project memory
Why a record, not a prompt
Retyping is where drift starts
A character pasted by hand into forty prompts is forty chances to describe them slightly differently. Held as one record, the description is written once and inherited — the shots cannot disagree about who this is.
Nothing leaves your machine
CastLock stores characters, versions and sequences locally in your browser. There is no account and no upload — the work stays on the machine you did it on.
Panels are for the model
A labelled collage sheet is useless as model input — handed a grid, the model tries to render the grid. Forge produces one subject, plain background, no text per panel, which is what a reference upload has to be. The bible sheet is composed in your browser from those same panels, for people.
Then prove it holds
A record guarantees the words are identical every time. The LoRA lab tests whether the model actually renders the same person from them — trained from the panels you already forged, checked while it is still cheap to find out.
Open Cast → Forge panels → LoRA lab →
Step Two

Build Your Cinematic Prompt

CinePrompt is your prompt architect. Choose from hundreds of cinematic parameters — camera, lens, lighting, color grade, texture — or translate any existing prompt across platforms instantly.

02
Tool
CinePrompt
Three modes: Build a prompt from scratch using cinematic dropdowns, paste an existing prompt and Translate it to any other platform automatically, or take the character you cast and generate a locked Multi-shot set so every shot matches the first. Supports 11 platforms.
CinePrompt · Build
PlatformKling 3.0 ▾
CameraARRI ALEXA ▾
Lens85mm T1.4 ▾
LightPractical rim ▾
GradeTeal-Orange ▾
"Close-up, ARRI ALEXA, 85mm T1.4, teal-orange, practical rim..."
Save to DVP Vault →
BUILD mode — what you do
  • Describe your scene or subject in the text field
  • Select platform, camera, lens, aperture, shot type
  • Choose lighting, color grade, film texture, and format
  • Click Generate — get a platform-optimized prompt
  • Shooting a sequence? Switch to Multi-shot, load the character from Cast, and generate a locked set of prompts across shots
  • Every prompt is saved to the History panel — compare any two with word-diff A/B view or restore earlier settings with one click
TRANSLATE mode — what you do
  • Paste any existing prompt — any platform, any format
  • Select your source platform
  • Check the target platforms you want
  • Click Translate — AI rewrites for each platform's syntax
Key insight
Every platform has a different prompt dialect. Kling wants comma-separated descriptors. Runway wants evocative sentences. Veo wants director instructions. Translate mode handles all of this automatically.
Supported platforms
Kling 3.0 Runway Gen-4.5 Veo 3.1 Luma Ray 3.2 Hailuo / MiniMax Seedance 2.5 Midjourney V8.2 Grok Imagine Nano Banana 2 Pika 2.5 Wan 3.0
Open CinePrompt →
Step Three

Save the Project Memory

DVP Vault is where the workflow becomes reusable. After preparing a shot, approving its image, and reviewing its video, save the prompt recipe, references, and character notes. CastLoop media remains downloadable from CastLoop rather than a durable Vault attachment.

03
Tool
DVP Vault
A project vault for recipes, references, and characters. Save winning prompts from CinePrompt and the characters you cast, track platform-specific IDs, and build a continuity record you can reopen later.
DVP Vault
📁 Project: Neon City
● Opening shot recipe
● Character: Sarah v2
● Style lock · teal-orange
● Ref Card · session 3
+ Save to vault
What you do
  • Save your strongest prompts from CinePrompt and the characters from Cast into a project vault
  • Attach characters, reference IDs, seeds, ratings, and notes; download approved CastLoop media separately
  • Keep multiple versions so you can compare iterations and avoid visual drift
  • When you save a recipe, DVP Vault checks whether the prompt drifts from the locked anchor it was generated against and flags any mismatches
  • Recall saved memory straight into VibeBoard or Campaign Engine when you start the next deliverable
  • Export the vault as portable Vault JSON when you want a backup or a copy for collaborators
How it should work in the flow
Not a starting tool
DVP Vault is the memory layer, not the ideation layer. Use it after you have built prompts, references, or consistency locks worth keeping.
Best use
Treat DVP Vault as the place you save approved looks, tested prompt recipes, character continuity notes, and reusable reference data for the whole production.
Open DVP Vault →
Phase 02 — Finishing

After Generation: Upscale & Deliver

After Prepare → Image → Video → Vault, the Finishing phase takes approved downloads to delivery quality. It does not touch your seeds, prompts, or continuity, so run it shot-by-shot or batch it after the creative work is locked.

04
Tool · Finishing
AI Upscaler
Take preview-quality Kling, Runway, Veo, or Midjourney output and finish it for delivery. Real-ESRGAN 4× upscaling for image and video, GFPGAN face restoration tuned for AI portraits, and RIFE / Bytedance frame interpolation to 48fps or 60fps. Runs on fal.ai with your own API key — pay per job, billed directly by fal.
AI Upscaler · fal.ai
▶ Video · 12.4s · 1920×1080
●Upscale 2× · Real-ESRGAN
●Face restore · GFPGAN
●Interpolate · 60fps
●Estimated cost · $0.42 – $0.96
◆ Upscale Video
What you do
  • Drop an image or video from your approved DVP Vault shots
  • Use Image Quick Presets (Portrait/Face, Anime, Quick 2×, Max 4×) to configure in one click, or dial in manually
  • Pick an upscale factor (2× or 4×) and optional face restore for portraits
  • For motion: boost framerate to 48fps or 60fps with RIFE / Bytedance
  • Queue multiple images in the Batch tab and run them all with one click — each saves to History automatically
  • Side-by-side compare slider, dynamic cost estimate per file, history of every job
How it should work in the flow
Post-production only
Upscaling is non-creative — it doesn't change composition, continuity, or seeds. Run it only on shots you've already approved in DVP Vault so you're not paying to enhance discarded takes.
Best use
For maximum quality on AI video: upscale first, then run the upscaled result through frame interpolation. Tests on long clips with tile: 0 at 4× can OOM — drop to 2× or use a shorter clip if the job fails.
Open AI Upscaler →
Quick Start

Your First Project in 15 Minutes

New to the suite? Follow this checklist. You'll go from zero to a full shot list with consistent prompts on your first run, then save the winning setup into DVP Vault.

1
Set up your API key
Get a free key at console.anthropic.com. Paste it into any tool — it saves automatically and works across the entire suite.
2
Decide what the shot is
Describe it in CinePrompt and set the lens, light and grade. Working from writing instead? Run the scene through Script to Shots first. Have a reference still? FrameDNA will read its cinematography and load it in.
3
Build or refine your prompt
In CinePrompt, refine the scene description and adjust camera and lighting settings. Shooting more than one shot of the same character? Use Multi-shot to generate a locked set.
4
Forge your cast and locations
In Cast's Forge mode, describe the character, prop or location and generate the hero image with your own Gemini key. Check it, then run the rest of the panels — you get clean reference images to upload to your video model, and a bible sheet for the room.
5
Cast the character
In Cast's Build room, your forged record is already in the library. Write the identity spine, mark the traits that must never drift, and save it as a version. In Sequence, attach it to your shot list so every shot inherits the same description.
6
Prove the look holds
In Cast's LoRA lab, train a Consistency Model from the panels you forged — no re-uploading — then compare and drift-check the results so you know the model renders the same person before you commit to a sequence.
7
Build your shot list
Paste your scene text into Script to Shots, or run CinePrompt's Multi-shot from the character you cast. Get one prompt per shot, per platform.
8
Save the approved recipe
Send your strongest prompt recipes, references, and character notes into DVP Vault so the project has a persistent memory you can reopen later.
9
Generate your shots
Copy prompts into Kling, Runway, Veo, or your platform of choice — or send the scene to a model straight from Cast. Your record and DVP Vault keep the visual style consistent across every generation session.
FAQ

Common Questions

Do I need to use every stop every time?
No. The pipeline is modular. If you already have a strong prompt, go straight to CinePrompt and shape it for your platform. If you're working from a script, start at Script to Shots. DVP Vault is where you save the approved work. Cast is what a character-led project runs on; for environment-only or one-off shots there is nothing to cast, so you can go around it.
Is my API key secure?
Yes. Your key is stored only in your browser's localStorage — it never touches DVP servers. You're billed directly by Anthropic for any Claude API usage (typically $0.01–$0.05 per generation). We have no access to your key or usage data.
Why do different platforms need different prompt formats?
Each AI model was trained differently. Kling responds to comma-separated descriptors and rewards camera movement first. Runway prefers evocative, period-separated sentences. Veo wants director-style instructions. Seedance anchors the subject in the first 20 words. Using the wrong format on any platform reduces quality significantly — that's why CinePrompt builds and translates platform-specifically.
What's the best platform for beginners?
Start with Kling 3.0 — it's the most forgiving with prompts and produces high-quality photorealistic video. Runway Gen-4.5 is excellent for stylized or VFX-heavy work. Veo 3.1 is best for physics-accurate motion and native audio. Try all three with the Translate tool to compare results from the same prompt.
How do I maintain consistency across sessions?
Keep the character as a record in Cast rather than as prompt text: every version is kept, the panels and any Consistency Model stay attached to it, and it is stored locally in your browser. Save the approved recipe to DVP Vault so the next session starts from the same identity, style settings, and platform IDs.
Where did FrameDNA go?
It is still here — it is just not a stage of the pipeline. FrameDNA reads the cinematography out of a reference image you already have, which is one way to fill in CinePrompt rather than a step every project takes. Use it when you have a still you want to work from; skip it when you are describing a shot from scratch. Same for Script to Shots at the front: useful when you are starting from writing, unnecessary when you already know the shot.
Why is Cast a step and not a support tool?
Because anything with a recurring character depends on it. Cast makes the look as images, turns it into a versioned record a whole shot list can inherit, and can train a Consistency Model to pin the face itself — that is the loop that keeps a cast recognizable from shot 1 to shot 40. For environment-only, abstract, or one-off shots there is nothing to cast, so you can go straight to CinePrompt.