Two finished avatar clips, two different outcomes

Both P-Video 2 Pro avatar jobs finished with a playable picture and an audio track. In the 480p take, the requested change appeared and the location changed with it. In the 768p take, the original styling survived and the requested change was softened. A log of completed jobs gives these two takes the same mark.

Three frames from a separate public sample. A fictional adult in a white jacket turns on a terrace. The stone railing and citrus tree remain in view. — Reading a shot at 0, 2 and 4 seconds. In this separate SFW demonstration, the terrace railing and citrus tree give the reviewer two fixed points while the performer moves. The two NSFW test files below stay private.

Both files passed a technical check

On 2 November 2026 the 480p run used about 5.21 seconds of speech. Its output contained 126 video frames at 24 fps. The 768p run used about 4.55 seconds of speech and returned 110 frames at the same rate. We measured the encoded audio and picture separately. Their end times in both files are within one frame of each other.

Figure 2. Each pair is aligned to the end of its own video track. The 480p file contains 5.213 seconds of audio and 5.250 of picture. The 768p file contains 4.550 seconds of audio and 4.583 of picture. These timestamps cannot establish whether a mouth movement matches a word.

Only the picture revealed which edits survived

Observations from the original visual review · one file per setup
Take	NSFW direction	Continuity note
A · 480p832 × 480	Requested change appeared.	The scene moved away from the source setting.
B · 768p1344 × 768	Requested change softened.	The original styling remained visible.

For a standalone character clip, take A may be usable. Cut it into a sequence that has already established the location and that same change becomes a continuity error. Replaying or remuxing take B will still show the original styling. Change the generation setup if that visual detail is essential to the shot.

Decide what must survive before the next run

On the review sheet, write one thing the avatar should change and one thing it must preserve. Watch once muted for those two details. Then play the spoken line with sound and inspect the first accented word and the end of the phrase. Keep separate entries for timing, the requested change and continuity, even if the take gets a single overall decision.

These runs changed the resolution and the audio length, and the second used prompt expansion. Neither file isolates the effect of expansion or resolution. To test expansion, keep the same reference, recording, size and brief across several pairs, then compare the original outputs with the setting on and off.

Open the talking-avatar workspace Score the next take →