Compositing Two Subjects Into One Frame

The constraint nobody reads the docs closely enough to catch

Some video-generation models accept exactly one conditioning image -- one start frame that anchors the whole animation. That's not a soft limit you can push against with a cleverer prompt. It's a hard interface constraint: the call has one slot for a start image, not two.

This becomes a real problem the moment a scene needs two independently-referenced subjects in the same shot -- a trained hero identity (see "Soul Cinema: Training an Identity Model for Exact Lookalikes") standing next to a separately-designed second character or creature that was never part of that identity training. You cannot hand the video model two separate reference images directly and expect it to honor both. Depending on the implementation, the call is rejected, or it silently only uses one of the two images and drops the other subject entirely -- which is worse, because nothing tells you it happened until you're looking at the output.

The workaround: solve it one step earlier

The fix isn't to fight the video model's interface. It's to stop asking the video model to solve a problem it was never built to solve, and push the multi-reference problem one step earlier in the pipeline -- onto an image-generation model that DOES accept multiple reference images.

The sequence:

Composite both subjects into one establishing frame. Call a multi-reference image model with both subject references. This is a single generated still that contains both subjects together, in the same frame, in roughly the composition the shot needs.

Review and lock the composite. Don't skip this. The composite frame is about to become the single source of truth for both identities for the rest of the shot -- if a face or a prop looks wrong here, it will look wrong in every frame of the animation that follows. Fix it now, while it's one cheap still, not after an expensive animation pass.

Feed the locked composite as the video model's single conditioning image. Now the video model has exactly what it wants: one start image. Both subject identities are already baked into it. The motion prompt drives the action from there, and both identities hold through the animation because they were never asked to ride in as two separate references -- they're one image now.

The ordering rule that actually matters

There's one detail in the composite step worth calling out on its own, because it's counterintuitive and it's a real lesson learned in production, not a guess: when composing a multi-reference prompt for the composite step, put the face/identity descriptor LAST in the reference ordering. Most people's instinct is to lead with the "most important" subject -- the trained hero identity -- but ordering it last is the pattern that reliably produces better results. Put the secondary character or object reference first, and close the reference list with the identity.

Why this pattern generalizes

The two-step composite-then-animate pattern isn't a one-off trick for a hero-plus-creature shot. It's the general answer any time a scene needs more independently-referenced subjects than the video model's single conditioning slot can hold -- two rival characters facing off, a trained identity opposite a second trained identity, a crowd shot with two named principals. The number of subjects doesn't change the fix: bind everyone who needs an independent reference into one composite frame first, lock it, then animate that one frame. The video model never has to know it's looking at a scene that started as two or three separate references -- by the time it sees anything, it's one image.

This is also the foundation the next lesson builds on directly. Twin and doppelganger shots -- two characters who deliberately share a face -- start from this same composite discipline, with one additional wrinkle: when both subjects are SUPPOSED to look alike, you stop fighting the identity model's face-mapping behavior and lean into it instead.

The cheap repair for the one-region fix

Even with a locked composite and a clean animation, a finished clip sometimes comes back with one small issue in one region of one frame -- a face that's slightly off in a single moment, nothing wrong with the rest of the shot. A full re-render to fix one region is expensive and often unnecessary. Some tools support redrawing a sketched region at a specific timestamp in an already-finished clip: a targeted repair pass, effectively a "poor-man's face-swap" for just the flawed region, instead of regenerating the entire shot from scratch. Reach for the targeted repair first; save the full re-render for when the composite itself was actually wrong.

Build it: fix the naive two-image call

You're going to replace a shot-builder function that naively tries to pass two independent subject references into one video-generation call with the correct two-step spec: composite first with the identity descriptor ordered last, then animate the single locked frame.