The LTX 2.3 Family: Fast, Pro, and Retake

LTX 2.3 is the first video model family where I stopped thinking in terms of "generate and pray" and started thinking in terms of drafts. Most models give you one lever: run the prompt again and hope. LTX 2.3 gives you three tools with different jobs — Fast for iteration, Pro for the final render, and Retake for surgically fixing the four seconds that went wrong without touching the eleven that didn't.

That last one is the quiet revolution. Before Retake, a 15-second clip with one bad segment meant regenerating the whole thing, re-rolling the dice on everything that was already working. Now I fix the segment. My regeneration spend on LTX projects dropped by roughly half the month I switched my workflow over, and my time-to-approved-clip dropped with it.

This is a working guide to the family: what each tier actually does, when to route to which, and the handoffs between them that make the whole thing feel like an editing suite instead of a slot machine.

The three tiers, in one table

Tier	Job	Native audio	Typical use	Relative cost
LTX 2.3 Fast	Rapid drafts from text	Yes	Prompt exploration, storyboards, social drafts	Lowest
LTX 2.3 Pro (I2V)	Final renders from an anchor image	Yes	Hero clips, client deliverables	Mid-high
LTX 2.3 Retake	Regenerate a segment of an existing video	Inherits	Fixing a bad beat without a full re-roll	Per-segment

The mental model: Fast is your sketchbook, Pro is your camera, Retake is your editor.

LTX 2.3 Fast: iterate before you commit

Fast is a text-to-video model tuned for turnaround. The output is genuinely usable — this is not a degraded preview mode — but its real value is that it makes prompt exploration cheap enough to do properly.

My loop on any new concept:

• Write 3–4 prompt variants of the same shot (different camera language, different pacing verbs).
• Run all of them on Fast.
• Pick the winner, extract or generate a matching anchor frame, and hand it to Pro.

Because Fast ships native audio, you also learn early whether your ambience and dialogue cues are landing. Finding out at draft stage that "distant traffic hum" reads as "aggressive motorway" costs you pennies. Finding out on a Pro render costs real money.

Multi-res matters here too. I draft vertical 9:16 for short-form and only occasionally re-run a 16:9 variant if the clip is destined for YouTube — the model handles both without the composition falling apart at the frame edges, which was not true of the 2.x builds before it.

LTX 2.3 Pro: the image-to-video finisher

Pro is where deliverables come from. It runs image-to-video, which means the input is an anchor frame you control — generated in a text-to-image model, pulled from brand photography, or extracted from a Fast draft you liked.

That anchor-frame step is the discipline most people skip. It buys you three things:

• Composition lock. The first frame is exactly what you approved, not the model's interpretation of your adjectives.
• Brand control. Product, palette, and framing come from an asset you can iterate on cheaply as a still.
• Series consistency. Reuse the same anchor (or variants of it) across a campaign and the clips feel related.

Pro's motion quality justifies the price gap on anything a client sees. Fabric, hair, liquid, and slow camera moves hold together at final-render scrutiny in a way Fast's drafts approximate but don't match.

Retake: the feature that changes the economics

Retake regenerates a chosen segment of an existing LTX video while leaving the rest untouched. In practice it's the difference between AI video as a lottery and AI video as an edit.

Where it earns its keep:

• One bad gesture. A hand glitches at second 9 of a 15-second clip. Retake seconds 8–11, keep everything else.
• A dead beat. The middle of the clip sags; you retake it with a stronger action cue in the prompt.
• Product fidelity. The label warps for one moment mid-move. Fix the moment.

Rules I've learned the hard way:

1. Retake slightly wider than the flaw. Grabbing a half-second on either side of the bad frames gives the model room to blend motion in and out.
2. Keep the prompt delta small. Retake with a nearly identical prompt plus one corrective clause. Rewriting the whole prompt for a segment invites a visible style seam.
3. Two retakes, then re-anchor. If a segment won't come right after two attempts, the problem is upstream — go back to Pro with a better anchor or prompt instead of grinding.

For context on how this compares economically to closed-source alternatives, the LTX 2.3 vs commercial models ROI breakdown runs the numbers; Retake is a big part of why the math favors LTX for iteration-heavy work.

Native audio: what it's good for, honestly

All current LTX 2.3 tiers generate audio with the video — ambience, effects, and dialogue-adjacent sound. My honest scorecard after a few months:

• Ambience and room tone: strong. Café murmur, rain, wind, machine hum — usable as-is in most social cuts.
• Synced effects: good. Footsteps, door closes, and impacts land close enough that I rarely foley over them for short-form.
• Dialogue: situational. Short lines work; anything script-critical I still generate separately with a proper TTS voice and lay over, especially for brand voiceover where I need a consistent narrator across a series.

The workflow win is that even "temp" native audio makes drafts feel finished, which makes client review dramatically easier. People judge silent clips unfairly.

A full pipeline example

A recent product teaser for a DTC kitchenware brand, start to finish:

1. Six prompt variants on Fast, 9:16 — total cost trivial, 20 minutes including thinking.
2. Winning composition re-created as a still with the product comped in, then two Pro renders from that anchor.
3. One Pro render had a wobble in the pour at the 7-second mark. One Retake pass on seconds 6–9 fixed it.
4. Captions and a music bed added, published to Reels and TikTok from the same project.

Three model calls that mattered, zero full-clip regenerations wasted. Under the old one-model workflow this would have been six or seven full renders to get an equivalent result.

If you're comparing LTX against the other open-weight contender, Wan 2.7 vs LTX covers that matchup; the short version is Wan wins on reference-driven character work, LTX wins on the draft-finish-fix loop.

Where the family still falls short

• Long single takes. Past the model's comfortable duration, coherence decays; plan multi-shot sequences instead of one heroic take.
• Dense text in frame. On-screen signage and packaging text still wobble under motion. Keep critical text to overlays.
• Retake can't change everything. It fixes motion and moments; it can't restyle the whole clip or change the subject mid-video. That's a regeneration, not a retake.

FAQ

What's the difference between LTX 2.3 Fast and Pro?

Fast is a text-to-video tier tuned for cheap, quick drafts; Pro is an image-to-video tier tuned for final quality from an anchor frame you control. The intended workflow is to explore on Fast, then finish on Pro — not to pick one and ignore the other.

What exactly does LTX 2.3 Retake do?

Retake regenerates a time segment of an existing LTX video while leaving the rest of the clip untouched. You select the flawed span, adjust the prompt minimally, and the model re-renders just that window and blends it back in. It's for fixing moments, not restyling whole videos.

Does LTX 2.3 generate audio with the video?

Yes — ambience, synced effects, and basic dialogue come out with the render across the family. Ambience and effects are usable in published work; for script-critical voiceover most teams still layer a dedicated TTS voice for consistency across a campaign.

Is LTX 2.3 good enough for client deliverables?

Pro renders are, for short-form and most web work. The practical ceiling is long single takes and legible in-frame text, which are weaknesses shared by most 2026 video models. Draft on Fast, deliver from Pro, and budget one Retake pass per hero clip.

All three LTX 2.3 tiers run inside Versely's AI video generator alongside 60+ other models, with live rankings on /models if you want to see where each tier sits this week. Free credits daily — draft something on Fast tonight and see how far the loop takes you.