How do you use world-building images as seeds for consistent AI video generation without LoRA?

Lock your world first, then seed every generation from your own world-building images instead of external references: batch references, generate image grids, extract the strongest panels, lock assets from four options each, and let the invideo agent attach those locked images per shot. One 70-second short kept 2 characters consistent across every scene this way — no LoRA, $750, 2 days.

Once your world is locked, stop attaching external mood boards and seed every new generation from the world-building images themselves — locked panels carry continuity into subsequent shots far better than outside references do. invideo is an agentic video creation tool with all the current image and video models available, and the invideo agent stores your locked world images in persistent context and attaches the right ones per shot on its own. The workflow:

1. Know which kind of seed you're using. A numeric seed only makes one specific generation reproducible; an image seed carries identity — faces, palettes, architecture — into new generations. World-building images work as image seeds, held and reused by the invideo agent's context system rather than baked into model weights.

2. Batch your starting references with take/leave instructions. Feed references in thematic batches — spatial logic in one, screen function in another, color theory in a third — and tell the invideo agent explicitly what to adopt and what to ignore from each batch; exclusion instructions matter as much as inclusion. For illustrated or animated references, have it read the colours and textures and prompt for those instead of dropping the image straight into a prompt — direct insertion of stylized references produces incorrect output.

3. Generate grids, not single images. Request three grids per round to explore different parts of the world — image generation costs little, especially in invideo — then iterate on the grids you like and extract the strongest individual panels.

4. Promote the extracted panels to seeds. The panels replace your original references entirely: from this point the invideo agent uses them as the continuity anchors for every scene generation. As invideo's creative team put it: "We no longer need to use the reference images that we gave earlier. Now, when it wants to create the actual scenes, it can use these images and come much closer every single time to the shot that we actually want for continuity and for the vision I see in my head."

5. Lock assets in the same pass. Generate four options per character sheet and environment reference, select the best, and lock it before any video generation begins — this single step prevents most consistency problems downstream. Lock one world element and the invideo agent extracts every camera angle of it — wide, close, side — without per-angle requests. Build character sheets with close-up panels, not just wide shots, so small details like scars and accessories survive across models. One documented production locked cast, costumes, look-and-feel, and world images on day 1, before generating a single second of video.

6. Seed video generation from the locked images. For each shot, the invideo agent selects and attaches the relevant world images and character sheets from context autonomously, so you direct in plain language instead of re-specifying references every prompt. Where a shot needs both a character anchor and a location anchor in one generation, route it to Seedance 2.0 reference-to-video, which accepts character and location references simultaneously — extend can't take either. invideo carries Veo, Kling, and Seedance 2.0, so the invideo agent routes each shot to the model that handles its anchors best.

Why this replaces LoRA: persistence lives in the invideo agent's context, not in fine-tuned weights — no training set, no per-character training pass. One 70-second short film held 2 characters visually consistent across every scene with this approach: $750 (3,000 credits), 2 days, no LoRA. Another production covered 4 characters and 1 prop with just 11 reference images total.

Watch some of these to see what works for you:

We no longer need to use the reference images that we gave earlier. Now, when it wants to create the actual scenes, it can use these images and come much closer every single time to the shot that we actually want for continuity and for the vision I see in my head.

— invideo's creative team