Reference Assets That Hold
Every complaint about AI video — the face that drifts, the room that swims, the logo that comes back subtly wrong — is the same complaint. The model was allowed to improvise something you had already decided. This lesson is about taking that decision away from it.
One idea, three rules
A video model fills in whatever you leave ambiguous. It is very good at this and that is the problem: it will confidently invent your character's jaw, the depth of your room, and which of your four props you meant. It will invent them differently every run.
So the entire craft of reference preparation is removing ambiguity before the model ever sees the shot. Three rules do most of the work.
The rules
1. Character sheets go on neutral gray
Build every character reference on flat neutral gray — around #777777. Gray outperforms white, black, and green.
Why: white and black sit at the ends of the exposure range, so the model compensates for the backdrop and drags your character's own values along with it. Green carries chroma-key association and bleeds into hair and edges. Neutral gray is mid-luminance and has no hue, so the model reads the character instead of the background.
Do: put several views on one sheet — full body, headshot, three-quarter turn. Ambiguity about a jawline or how a jacket drapes is ambiguity the model resolves for you, and it will resolve it differently next time.
2. Location plates go on a three-quarter angle
Never hand the model a flat, straight-on photograph of a space. Frame it from a three-quarter angle so two surfaces meet at a corner.
Why: a flat elevation has no depth cues — no parallax, no receding lines, no read on how far the back wall sits from the front of the room. Move a camera through that and the model has to invent the third dimension mid-shot, which is what "the geometry swims" actually is. A corner tells it the volume of the space in one frame.
Do: keep the ground plane visible so it knows where the floor is, and let objects overlap at different depths — occlusion is a depth cue the model uses.
3. Props need light, not outlines
A prop reference wants studio lighting — highlights on the bevels, a taper on the edges, a reflection that tells you which way the surface turns.
Why: hand the model a flat, evenly-lit cutout and it animates a piece of cardboard, because a flat image genuinely is one. Shading is how it learns the object has a back.
Name assets so you can find them again
An approved asset is only worth having if you can retrieve it six weeks later — and if an automated pipeline can retrieve it without picking the wrong one. Give every reusable asset a type prefix, a short project code, and an underscore-joined description:
loc_RR_suburban_porch char_RR_contractor prop_RR_yard_sign
Three prefixes cover nearly everything — loc_, char_, prop_. The project code keeps two productions from fighting over "front door." Agree it at kickoff; renaming later breaks every reference.
Why ours is a folder, not a picker
Cloud tools need a naming convention because your assets live in someone else's web library and the name is how you scroll to it. Yours live on your own disk, which means the convention can be enforced by the filesystem itself:
assets/<PROJECT>/char/contractor.png assets/<PROJECT>/loc/suburban_porch.png assets/<PROJECT>/prop/yard_sign.png
A folder structure is a naming convention nobody can forget to follow. And because the files are local, typing @ in any Hybrig prompt resolves straight to a path at render time — no upload step, no picker, no waiting on a server to hand your own artwork back to you.
Where this ends up
Push this principle all the way and you get the render farm. Solve the shot once — camera, light, motion, every frame — and the model is not permitted to improvise anything at all. It swaps one logo onto one tracked surface, and every customer gets a commercial that is identical except for the part that is theirs.
That is the same rule as the gray background, taken to its conclusion. See the render farm for what it looks like at 72 brands, and clean-plate compositing for why the personalization happens at composition time rather than at generation time.
Prove it yourself
Two of the three rules above are received wisdom. They are widely taught and we have not measured them here.
Everywhere else you would have to take that on faith, because verifying it costs money — every A/B burns credits, so nobody runs the test and the advice gets passed along on authority. Locally it costs electricity. So run it: same seed, same character, gray versus white versus black versus green, ten generations each. Lay them side by side and decide for yourself.
A lesson here is not "proven" until there is a contact sheet next to it. If your test disagrees with this page, the page is what changes.