For thirty five episodes our presenter Xellia was a photograph with a mouth. In EP36 she stands up, crosses four different rooms, and hands off between stations in what reads as one continuous take. Everything ran locally on one RTX 5090, on ComfyUI, with LTX 2.5. This guide is the full recipe, including the parts that failed, because the failures are where the settings come from.

The stack, named

  • ComfyUI with the built-in template LTX-2.5 First/Last Frame to Video (video_ltx2_5_flf2v)
  • LTX 2.5 (open weights) for every move and demo
  • Plates: CyberRealistic Krea 2 v10 + our character LoRA (epoch 20), 8 steps, CFG 1.0
  • Identity in motion: our LTX character LoRA (rank 32), strength 0.7 to 1.0
  • Lip-synced presenter passes: LongCat + MuseTalk, 832x480, exactly as in our earlier episodes

The architecture: talk, move, talk

She talks at a station. The move starts from the last rendered frame of that talking clip and ends on the next station’s plate, which is also that station’s first frame. Every join is frame-exact, so no crossfades are needed, and speech pauses during moves so lip-sync never touches them. Half of the model’s usual disagreement is gone before it starts, because one endpoint of every move is not a guess.

Technique 1: the walk-set (one prop per frame)

Do not try to hold one room consistent across plates. The image model has no 3D scene, so it will keep the furniture and shuffle the layout, and the video model then orbits the room trying to reconcile it. Instead: each plate contains exactly one prop, she exits left, the camera pans left with her. A lateral track is expected to reveal new background, so the background changing reads as camera movement. Continuity rides on her: same person, same clothes, same light.

Two hard-won details:

  • Negate the other props by name. Asking for one prop is not enough; the LoRA pulls the rest back in.
  • Show the edge of the next prop entering at frame left. Otherwise the destination prop has to be invented mid-move and arrives as a smear. But this is a narrow rule, not doctrine: if the destination already contains a prop she must interact with, do not pin extra objects into it.

Technique 2: the two settings that stopped the dissolves

The template ships LTXVAddGuide at 0.7 (not 1.0, so the usual “lower it from 1.0” advice would raise ours). Dropping both guide nodes to 0.5 is what stopped our cross-fades. Holding the endpoints less tightly gives the model room to find an actual path between them. We isolated this properly on our own footage: same plates, same seed, no other changes; at 0.7 she goes semi-transparent mid-move, at 0.5 the same move is dramatically cleaner. Endpoint adherence did not measurably suffer.

And the honest half: our negative-prompt list (cross dissolve, double exposure, ghosting, morphing, fading, frame blending, transparent body, duplicated subject) stays in our config, but when we went looking for the clip that proves it does anything on its own, we could not make one. Same pair, negatives off and on, both guide settings, three seeds, two sets: identical frames every time. We teach what we measured, so: the guide strength is the fix; the negatives are a habit we have not been able to isolate. If you find a case where they clearly help, we want to see it.

Technique 3: the middle is the fragile part

Both endpoints are hard conditioning; the middle is least constrained. Measuring each frame’s distance to both endpoints puts the least-anchored moment at roughly 47% through the clip, and that is where artifacts land. Changing what is in frame changes what ghosts, not whether something ghosts. For a hard move, generate in halves around a mid anchor lifted from a first-pass render.

Technique 4: generate long, cut short

The model reaches its destination early and then holds the final pose. The hold is roughly constant, not proportional to length. So ask for more seconds than you need and trim the dead hold, splicing inside the still section so the plate-matched final frame survives. Duration should match the distance of the move: too little time forces motion blur, too much invites wandering.

Technique 5: when a lateral move is impossible

If the same large object sits on the same side of both endpoints, no walk can resolve the pair: she cannot travel away from a camera that must still be there at the end, and the model wanders in circles instead. The escape is to stop asking for a traverse. Have her turn in place, take the prop, let the camera move in close, then pull back out to the new arrangement.

What we learned prompting it, the expensive way:

  • Direct like a director, not a camera operator. “Zooming in just enough to see her smile” works; “zoom in until her head completely fills the frame” gets faked as a defocus blur with a crossfade hiding behind it.
  • Keep the prop visibly in her hands through the move and object-teleport defects disappear.
  • Do not name what should be unseen. “Lose sight of the floor” made the model zoom at the floor and furnish it. Omitting the floor entirely produced clean takes.
  • Capital letters mean nothing to diffusion models.
  • Emotion direction lands: “smile” produced her most alive take, and she blinks naturally unprompted. A deliberate wink never landed in any seed, on stills or video. Some gestures still need a reference, not a sentence.

Technique 6: identity under the close-up

In a normal move the plates carry her face and the video model only interpolates. The moment the camera comes close, the video model is authoring her face at scale, and without help she drifts generic. Load the character video LoRA (ours at 0.7 to 1.0) and put the trigger in the prompt. At 1.0 identity is strongest but the final frame can drift off the plate on some seeds, loosening the join; check the last frame against the plate whenever you raise it.

Technique 7: the green screen, honestly labelled

Generate both endpoints with a flat chroma-green screen in frame, run the move, key the green, and put a real rendered graphic behind her. The key tolerance that worked on our footage was 0.28 (0.16 keyed nothing, 0.22 left the bright centre, 0.32 started eating her), because the model renders a lit desaturated green rather than pure chroma, and you key against pure green. Sweep it on your own footage rather than trusting ours. We are calling our chroma work an early experiment: screen-angle matching and edge locking are not solved here yet, and we will publish a proper guide when they are.

What it costs, on our card

On one RTX 5090 with our workflow, a five-second move on a sparse set took about a minute; the same move on a prop-heavy set took two to three minutes; the single-ended method we started with took about twenty nine minutes for eight seconds. Those are our numbers on our card, with another ComfyUI instance sharing it. Yours will differ; the ratio is the part that should travel.

The QC that saved this episode

Six frames of a five-second clip is under 5% of it; defects live between your samples. Inspect at 4fps minimum, and inspect the whole clip, because the next defect never appears where the last one did. Scores shortlist, eyes decide: our endpoint metric once ranked a take where the presenter had walked out of frame as the best of the batch, because an empty set matches an empty plate beautifully.

Everything here ran on the same machine that renders our episodes. The video walks through all of it with the failed takes on screen, because the failures are the instructions.