DEV Community: Xinyue Zhan

My AI video toolchain: four steps, three tools, in the order that works

Xinyue Zhan — Tue, 10 Nov 2026 08:08:06 +0000

The tool list is short. What took me months to get right was the ordering. Here is the chain I run on a shoot day, step by step, with the one job each tool is there for.

Step 1: storyboard in a text file, not in a tool

Six lines, one per shot: subject, framing, light, duration.

hands opening a box, close-up, warm rim light, 4s
product on stone surface, medium, soft window light, 3s
...

No tool touches this step. It is also the step that decides whether everything after it is fast or slow — every one of the tools below produces better output from a concrete shot description than from a mood word.

Step 2: one-beat shots with Wan 3.0

Wan 3.0 does text-to-video and image-to-video with motion presets. It is the first tool in the chain because it covers the shots that do not need continuity: a push-in on hands, a slow reveal of a product, a background plate.

Two habits that matter:

Pick the motion preset before editing the prompt. A slow push-in and a handheld drift read completely differently, and the preset is the control that actually changes the feel.
Re-roll instead of editing. The shot is short; regenerating costs less than fiddling.

Step 3: the sequence — cut, caption, dub — in Pavo AI

Pavo AI is where the project becomes a video. Generation, timeline editing, captions and multilingual dubbing share one workspace, which removed the export/import dance I used to do between three tools.

The order inside this step matters as much as the step itself:

Drop every shot on one timeline — 9:16 master.
Cut to the storyboard rhythm. This is where you find out two shots are missing.
Character consistency: lock the reference frame once and reuse it for every generated shot. That single constraint removed most of my rework — without it, shot three drifts from shot one and you end up regenerating the sequence.
Captions next, because they change where the eye lands.
Dub last. Dubbing before the cut is locked means re-dubbing every trim.
Export 16:9 from the same timeline for the landing page.

Prompt examples for the locked-reference approach: pavoai.app/prompts.

Step 4: match an existing reference with Rea Video

Sometimes a shot has to look like something that already exists — a competitor's ad, a client's old spot, a still the client loves. Generating from a description is the wrong move here; you want recreation.

Rea Video is reference-driven: give it the clip, describe the new subject, and it re-shoots the scene. Describe only the subject change — if you also describe staging, you are asking it to invent, and it drifts.

Why the order matters more than the tools

Generating before the storyboard wastes credits on shots you will cut.
Dubbing before the cut is locked wastes a whole pass.
Re-shooting something a preset could have generated wastes time.
Skipping the character reference lock guarantees a regeneration round.

The chain, one table

Step	Tool	The one job
Storyboard	a text file	six shot lines
One-beat shots	Wan 3.0	motion presets, no timeline
Sequence, cut, caption, dub	Pavo AI	everything that must hang together
Match a reference	Rea Video	re-shoot instead of generate

Chain links: wan-3.io · pavoai.app · reavideo.pro

Reverse-engineering a reference video: a Rea Video walkthrough

Xinyue Zhan — Tue, 10 Nov 2026 07:56:56 +0000

I had a competitor clip that worked and no idea how to reproduce the look. This is the process I used, in order, with the tool that did each step.

Step 0: describe the reference before touching a tool

Watch the clip twice on mute. Write down, in this order:

Shot count and cut rhythm — one shot or five? Cuts on beat or on motion?
Framing per shot — wide, medium, close. Where is the subject in frame?
Camera behaviour — static, push-in, drift, handheld. This is the one people skip, and it is the one that decides which tool to use.
Light — direction, hardness, colour temperature.

If you cannot describe the shot, no tool will reproduce it. The description is the input.

Step 1: re-shoot the shot with Rea Video

Rea Video is reference-driven: hand it the clip, describe the new subject, and it re-creates the shot with your subject in it. Two things I learned:

Feed a reference where the camera work is legible. Fast cutting confuses the recreation; a clean 3-second push-in does not.
In the description, change only the subject. If you also describe staging, you are asking it to invent, and it will drift.

The output is a first pass. Expect to re-roll two or three times per shot.

Step 2: fix the single-shot gaps with Wan 3.0

Some shots in the reference are one clean beat — a push-in on hands, a slow reveal. For those, re-shooting is slower than generating. Wan 3.0 does text-to-video and image-to-video with motion presets, and the preset is the useful control: pick slow push-in and the clip reads right before you touch the prompt.

Where I used it instead of Rea Video:

Shots with no subject to swap (backgrounds, transitions, cutaways).
Shots where the reference was too low-res to re-create cleanly.

Step 3: assemble, caption and dub in Pavo AI

Once the shots exist, you have a sequence, and a sequence needs a timeline. Pavo AI covers generation, timeline editing, captions and multilingual dubbing in one workspace. The workflow that finally worked:

Drop all shots on one timeline, 9:16 master.
Cut to the reference rhythm — this is where you find out you generated two shots too many.
Captions next, because they change where the eye lands.
Dub last. Dubbing before the cut is locked means re-dubbing every time you trim a shot.
Export 16:9 from the same timeline for the landing page.

Character consistency is the part that makes this viable: lock the reference frame once, reuse it for every shot, and shot three still matches shot one. Without that lock I was re-generating whole sequences.

Prompt examples for the locked-reference approach: pavoai.app/prompts.

What the split actually looks like

Step	Tool	Why this one
Re-create a shot from a reference	Rea Video	it re-shoots instead of generating from scratch
Generate a missing single shot	Wan 3.0	motion presets, no timeline overhead
Cut, caption, dub the sequence	Pavo AI	generation and edit share one tab

The honest summary

Most of the work is step 0. The tools are fast; the description is slow. If you skip the shot description and start generating, you will burn an hour re-rolling and end up writing the description anyway.

Links: reavideo.pro · wan-3.io · pavoai.app

Which AI video tool for which job: a decision guide (Wan 3.0, Pavo AI, Rea Video)

Xinyue Zhan — Tue, 10 Nov 2026 06:47:51 +0000

Most AI video tool articles list ten options and rank none of them. This one is a decision guide: three tools, the exact job each one is wrong for, and the two-line test I use to pick.

The test I actually run

Two questions, in order:

Do I already have a reference? If yes, stop — you want re-shooting, not generation.
Is this one shot or a sequence? One shot goes to a generator; a sequence needs an editor and a dubbing step.

That splits the work cleanly. All three tools below are browser-based, none needs an install, and none of them is a general-purpose NLE.

One shot, camera move is the point: Wan 3.0

Wan 3.0 does text-to-video and image-to-video with motion presets. The presets are the product: a slow push-in versus a handheld drift changes the read of a clip more than any prompt tweak.

Wrong for: anything longer than a few shots. There is no timeline to speak of, and that is fine.

Site: wan-3.io

A sequence that needs editing and dubbing: Pavo AI

Pavo AI is the one I keep open for real work, because generation, timeline editing, captions and multilingual dubbing live in the same tab. The feature that decided it for me was character consistency: lock the reference frame once, and shot three still looks like shot one.

The loop:

Storyboard six shots as text.
Generate each with a locked character reference.
Cut the sequence.
Dub last, never before the cut is locked.
Export 9:16 and 16:9 from the same timeline.

Wrong for: a single 5s clip. You will spend more time in the timeline than generating.

Site: pavoai.app - prompts: pavoai.app/prompts

An existing clip needs a variant: Rea Video

Rea Video is reference-driven. Give it a clip you like, describe the new subject, and it re-creates the shot with your subject in it.

Wrong for: original creative. It is a re-shooter, not a brainstorm.

Site: reavideo.pro

The table

Situation	Tool	Why
One shot, specific camera move	Wan 3.0	motion presets, no timeline overhead
Sequence, needs edit + dub	Pavo AI	generation, editing and dubbing in one tab
Have a reference clip	Rea Video	re-shoots instead of generating from scratch

The thing that matters more than the tool

Write the shot list first. Every one of these produces better output from a concrete shot description than from a mood word. The tool choice matters less than whether the input is a shot description.

Links: wan-3.io - pavoai.app - reavideo.pro

Comparing three AI video tools: Wan 3.0, Pavo AI and Rea Video

Xinyue Zhan — Mon, 09 Nov 2026 12:28:50 +0000

I ran the same short-clip brief through three browser-based AI video tools.

Wan 3.0 - fastest path to a clip

Wan 3.0 is a text-to-video and image-to-video workspace with motion presets. Drop a prompt, pick a camera move, get a 5-10s clip.

Pavo AI - the whole loop in one place

Pavo AI covers generation, timeline editing, captions and multilingual dubbing. Lock the reference frame once and each shot keeps the same subject.

Storyboard six shots as text.
Generate each shot with a locked character reference.
Cut first, dub second.
Export 9:16 and 16:9 from one timeline.

Prompts: pavoai.app/prompts - workspace: pavoai.app.

Rea Video - re-shoot what already works

Rea Video is reference-driven: hand it a clip, describe the new subject, it re-creates the shot.

Side by side

Need	Tool
Quick clip, motion presets	Wan 3.0
Storyboard, generate, edit, dub	Pavo AI
Re-shoot a reference clip	Rea Video

Links: wan-3.io - pavoai.app - reavideo.pro

A practical AI video workflow: storyboard, generate, dub (with Pavo AI)

Xinyue Zhan — Fri, 06 Nov 2026 10:44:15 +0000

When you ship short-form video every week, the slow part is not the idea — it is the loop between storyboard, generation, and dubbing. I have been testing a browser-based workflow around Pavo AI that keeps all three steps in one place.

The loop

Storyboard — write 6 shots as text, one line each.
Generate — turn each line into a 4-8s clip, keep the same character reference across shots.
Dub — upload the cut, pick a target language, re-sync the lips.
Export — 9:16 for shorts, 16:9 for landing pages.

What actually helped

Keeping the character reference consistent between image-to-video calls was the main quality win. Regenerating a single shot instead of the whole sequence cut my rework time roughly in half.

The tool I used for the full loop is Pavo AI — text-to-image, image-to-video, and dubbing in one workspace, no install.

Takeaways

Write the shot list before touching any model.
Lock the character reference early.
Dub last, never before the cut is locked.
Keep a per-shot changelog so you can re-roll one shot without losing the rest.

If you want to try the same loop: https://pavoai.app