How to Build an AI Shot Generation Workflow for Pre-Production

The traditional pre-production workflow has a gap. You write a script, break it into a shot list, sketch a storyboard — and then you wait. You wait for a storyboard artist, or you sketch rough thumbnails yourself, or you pull reference images from a mood board that only loosely captures what you're after.

AI image and video generation closes that gap. Not by replacing creative direction, but by giving it a visual form before you've touched a camera.

This guide walks through how to set up an AI shot generation workflow for pre-production — from the first storyboard entry to having reference visuals your DP, client, and director can actually react to.

What AI Shot Generation Is (and Isn't)

AI shot generation produces reference visuals from text prompts. The output is not your final shot — it's a production reference. Think of it the way a DP uses a still camera to scout locations before the shoot: the frame isn't the film, but it tells you where the light is, how the lens behaves, and whether your vision holds up in reality.

Where AI generation adds value in pre-production:

Client approval before production — show a client what the visual tone looks like before committing to a shoot day
Director-DP alignment — agree on framing, color grade, and lens character with visual references rather than verbal descriptions
Shot-by-shot iteration — test 10 variations of a single shot concept in the time it would take to sketch one

Where it doesn't replace human work:

Precise set design, costume, or talent-specific visuals
Final color grade decisions
Anything requiring real-world lighting references

Setting Up a Shot Generation Pipeline

A production-grade AI shot generation pipeline has three stages:

Stage 1: Structured Shot Inputs

Start with your shot list, not an open prompt box. Each shot should have:

Shot type (wide, medium, close-up, OTS, POV)
Subject (who or what is in frame)
Action (what is happening)
Camera movement (static, dolly, handheld)
Mood / look reference (a film, a photographer, a color palette)

In KroxFlow's Storyline module, each shot card captures these fields directly. The structured data then feeds the Shot Generation Canvas rather than starting from a blank prompt.

Stage 2: Prompt Engineering for Cinematic Output

The difference between a generic AI image and a useful production reference is prompt structure. Use this framework:

[Shot type], [subject + action], [lens + camera], [lighting], [mood/style], [technical suffix]

Example for a golden hour exterior:

Wide shot, woman walking along empty coastline, 35mm lens, shallow depth of field,
golden hour backlight with lens flare, melancholic mood, cinematic, film grain,
8K, photorealistic, no text, no watermark

Key principles:

Lead with the shot type — models respond better when the framing instruction comes first
Name your lens — "35mm" vs "85mm" vs "14mm" produces visually distinct results in most image models
Specify lighting explicitly — "golden hour backlight" is better than "nice lighting"
Add negative prompts — "no watermark, no text, no cartoon, no anime" prevent off-brand outputs

Stage 3: Model Selection Per Shot

Not every shot needs the same model. A rough blocking reference for a client deck needs fast and cheap. A hero shot reference that the director will scrutinize needs high quality.

Model tiers for cinematic references (on FAL.ai):

Fast iteration: Flux Schnell — 4–6 steps, low cost, good for rapid concept exploration
Production quality: Flux Pro / Stable Diffusion 3.5 Large — higher fidelity, better prompt adherence
Video references: Kling, Runway Gen-3, Luma Dream Machine — for shots where motion matters

In KroxFlow's Shot Generation Canvas, you pick the model per node. A single storyboard shot can have a fast-iteration node for early rounds and a quality node for final client-facing references, all in the same graph.

Connecting AI References to Your Editing Workflow

The gap in most AI generation workflows is export. You generate in Midjourney, download to your desktop, manually organize into a folder, share via email or Frame.io, and hope the naming convention survives.

A better approach: keep references inside your storyboard.

In KroxFlow, each generated image is saved to the shot that produced it. When you upload your first cut to Media Review, editors and directors can pull up the storyboard in a split view and compare the AI reference against the actual footage. Shot-by-shot alignment, no separate reference library to manage.

Practical Tips for Pre-Production AI Workflows

Generate before you brief the DP. Come to your DP tech scout with visual references, not just a look book. It cuts the "what do you mean by moody?" conversation in half.

Generate multiple aspect ratios. Most image models generate square or widescreen by default. For storyboards, 16:9 is usually right. For client decks, square sometimes reads better. Specify --ar 16:9 or the equivalent parameter for your model.

Name your shots before you generate. "Shot_023_GH_Exterior_Wide" is a recoverable filename. "IMG_4782" is not. Establish naming conventions in your shot list before generation runs so you're not reorganizing 200 images afterward.

Use upscalers for hero frames. If a reference shot is going on a pitch deck or client presentation, run it through an upscaler (FAL.ai has several) after generation. The quality difference at presentation size is significant.

Regenerate with director notes. Treat each generation round like a revision round. Collect director notes in writing, translate them into prompt changes, regenerate. Three rounds is usually enough to lock a shot's visual direction.

KroxFlow's Shot Generation Canvas is available in beta. Join the waitlist to get access to the full pre-production workflow.