How do you use a mood board to guide and anchor AI image generation before building out campaign scenes?

Generate a 6–9 image hero shot mood board first — a grid showing the product, character, or world in varied angles, lighting, palettes, and rendering moods — review and lock it, then store it in the invideo agent's context so every campaign scene that follows inherits the locked visual language without re-prompting.

Start by briefing the invideo agent — an agentic video tool that holds project memory across every generation — to produce a mood board grid before any scene work. Ask for 9 images in one pass covering the angles, lighting setups, and compositions you want to explore for the campaign's hero subject (product, character, or location). This is the visual anchor: review it like a creative director, kill the frames that drift, and lock the ones that define the look.

Label each locked frame by visual dimension. Don't treat the board as a vibe collage — tag each image by what it contributes: palette, lighting direction, rendering style, spatial depth, composition, mood. This is the step that converts the board into prompt-ready inputs. When you brief the next generation, pull attributes individually — pose from one frame, lighting from another, framing from a third — and tell the agent that explicitly in chat ("I want the depth structure from image 3, the warm key light from image 7, the framing from image 1"). Decomposed reference pulling beats dumping all nine images as a single composite, which usually blends into mush.

Write the board's rules into the context tab. Translate the locked dimensions into a short written visual code the agent stores once and applies everywhere: lighting direction ("single hard warm key light, side-raked, environment falling into shadow"), depth structure (foreground prop plane, subject plane, midground transition objects, painted backdrop), palette, posing register, and a short list of standing don'ts (no plastic-looking skin, no generic AI faces, no garments that fight the backdrop). Hridaye, invideo's creative director, puts it directly: "Directing AI with craft-specific language — 'depth structure,' 'raked light' — produces more accurate results than generic similarity prompts."

Probe before you batch. Generate ONE full-frame shot containing every element the campaign needs to hold — character, garment or product, set, skin, pose, lighting — and check it against the mood board. If that probe holds, the system is ready to scale to the full shot list; if it breaks, fix the rule in context before committing credits. Skipping this propagates the same failure across every scene.

Then expand the locked board into a hero shot grid for the campaign. Ask the agent for a 3x3 coverage sheet that translates the mood board's visual code into actual campaign beats — close, mid, wide; product still life; subject in environment; back shot; atmospheric — and run an editorial gap audit on the list ("what's missing for a proper editorial?"). This surfaces pause beats the standard brief would skip: product still lifes, cropped fragments, empty atmospherics, duo compositions. Lock that grid. From here, the agent routes each beat to the right model — GPT-Image-2 for realistic locations and text-bearing frames, Nano Banana for product lock-in on intricate hero items, Recraft for skin and casting portraits — without you having to pick. Every campaign scene generated from here inherits the mood board's lighting, palette, and depth structure because the rules live in project context.

One documented editorial campaign run this way produced 40 stills and 30 motion clips across 2 models and 5 locations in 3–4 hours, 630 credits ($150) including every rejected generation. Two ad-campaign productions came in at $33 (155 credits) and ~$195 respectively for full multi-shot delivery — variance is normal across team and scope; the mood-board lock is what makes any of those numbers repeatable.

Beyond the board itself: when you move into video, feed each locked still as a keyframe reference into Seedance 2.0 or Kling, and the visual language carries from board → image → motion without redrifting.

Watch some of these to see what works for you:

Directing AI with craft-specific language ('depth structure,' 'raked light') produces more accurate results than generic similarity prompts ('make it like this')

— Hridaye, invideo's creative director