Everyone shows you the perfect AI video they generated. Nobody talks about the 47 failed attempts before it. Lightricks just shipped the fix — and it’s open source.
The frustration every AI video creator knows
You’ve generated the perfect AI video. The character is right. The lighting is right. The pacing is right.
Except for two seconds in the middle. The dialogue lands wrong. Or the ending feels flat. Or the character’s expression in one shot doesn’t match the rest.
So you regenerate. And you lose everything that worked.
That’s the slot-machine problem of AI video generation — every regeneration is a fresh roll of the dice. You don’t get to keep the good parts. You don’t get to direct. You get whatever the model gives you.
LTX-2 Retake fixes that.
What Retake actually does
LTX-2 Retake lets you regenerate any 2-16 second segment of an existing video while keeping everything else identical. Same character. Same lighting. Same setting. Same voice. Only the moment you choose changes.
It’s section-level surgical regeneration. The model “remembers” everything around the segment you’re editing, generates new content for just that window, and stitches it back into the original clip.
This isn’t trimming. It isn’t style transfer. It isn’t a remix. It’s the model directing a new take of one specific moment in a scene it already understands.
Why this is different from everything else
The closest comparison most people reach for is Sora 2 Remix. They’re not the same thing.
| Feature | Sora 2 Remix | LTX-2 Retake |
|---|---|---|
| Preserves original character | ❌ Often changes face | ✅ Stays consistent |
| Preserves voice | ❌ Regenerates audio | ✅ Same voice & timbre |
| Preserves background | ❌ New environment | ✅ Identical |
| Surgical segment control | ❌ Regenerates whole clip | ✅ 2-16s window |
| Open source | ❌ Proprietary | ✅ Apache 2.0 |
Remix creates a related video. Retake edits this video.
The real-world use cases
Once you understand what Retake actually preserves, the use cases open up fast:
- Fix dialogue without re-shooting the take. Your character says the wrong line at second 9? Retake just seconds 8-12. Everything else stays.
- Test five CTAs from one base video. Same hero shot, five different endings. Brand A/B testing got 10x faster.
- Change emotional beats. The character looked happy when she should have looked contemplative? Retake the closing moment.
- Try alternate endings. Generate four versions of the last three seconds. Pick the best.
- Adjust pacing. Section feels rushed? Retake it slower.
The shift is psychological as much as technical: you stop thinking of AI generation as gambling and start thinking of it as directing.
The numbers that matter
| Spec | Value |
|---|---|
| Max video length | 20 seconds |
| Max retake segment | 16 seconds |
| Min retake segment | 2 seconds |
| File size limit | 100 MB |
| Cost | $0.10 per second of the edited portion only |
| License | Apache 2.0 |
| Available via | LTX Studio, fal.ai API, ComfyUI |
The pricing is the part most people miss. You’re not paying to regenerate the whole video — only the seconds you retake. A 4-second fix costs $0.40. Iterate aggressively without burning credits.
The #1 prompting mistake everyone makes
This part matters more than the rest of this article. Get prompting wrong and Retake feels broken. Get it right and it feels like magic.
Wrong (editing language):
- ❌ “Change the dialogue to…”
- ❌ “Fix the ending so…”
- ❌ “Replace the emotion with…”
Right (scene description):
- ✅ “The man says ‘this is perfect’ and smiles”
- ✅ “She looks away, contemplating”
- ✅ “The car pulls into the driveway as rain starts”
Retake isn’t a trim tool or a style transfer. It regenerates that specific segment with new direction, while the model remembers the context around it. Using editing language (“change,” “fix,” “replace”) confuses the model because it has no concept of the previous version — it only knows what the rest of the video looks like and what you want this moment to be.
Describe what happens. Don’t describe what to change.
The workflow
Once your video exists:
- Generate your base video (any LTX-2 mode — text-to-video, image-to-video, audio-to-video)
- Select the segment that needs work (e.g., seconds 8-12)
- Write a fresh scene description for what should happen in that window
- Hit Generate
That’s the entire flow. No timeline editing. No mask painting. No reference frames required.
Honest limitations
A few things Retake can’t do:
- Can’t retake the full 20 seconds. Segment cap is 16 seconds. If you need to retake everything, just regenerate from scratch — Retake is for surgical fixes, not full rewrites.
- Can’t extend the video. Use the separate
extend-videoendpoint for that. - Won’t work miracles on incoherent base footage. If your original generation has serious problems throughout, Retake won’t save it. It works best on videos that are 80-95% right.
- Subject to LTX-2’s broader limitations. Currently 1080p (with 4K available on the newer LTX-2.3), audio quality is good but not perfect, character consistency over very long sequences still drifts.
Why nobody’s talking about this yet
Three reasons:
- It’s new. LTX-2 launched in October 2025; the Retake pipeline is part of that family. Most creators are still in the “regenerate and pray” mental model from earlier generations of video tools.
- It’s open source. Open-source AI rarely gets the marketing budget of proprietary tools, so it spreads through builder networks first and creator networks second.
- The mental model is different. Until you’ve actually used it once, it’s hard to grasp why “regenerate 4 seconds while keeping everything else” is fundamentally different from existing tools.
Early adopter advantage on this one is real.
What’s actually shipped
The Retake pipeline is included in the official Lightricks/LTX-2 Python package. You can run it locally if you have an NVIDIA GPU, use it through ComfyUI’s built-in LTX-2 nodes, hit it via the fal.ai API at $0.10/second, or work with it visually in LTX Studio.
License is Apache 2.0, which means commercial use is permitted without restrictions for companies under $10M annual revenue. Larger commercial deployments need a license from Lightricks directly. For the vast majority of creators, builders, and small agencies, this is effectively free.
Get started
- GitHub: github.com/Lightricks/LTX-2
- ComfyUI integration: github.com/Lightricks/ComfyUI-LTXVideo
- Hosted API: fal.ai/models/fal-ai/ltx-2.3/retake-video
- Desktop app: github.com/Lightricks/LTX-Desktop
If you’re integrating it, look for RetakePipeline in the ltx-pipelines package. The pipeline takes your existing video, a start/end timestamp for the segment, and a prompt describing what happens in that window.
The bigger shift
AI video tools have been slot machines. You pull the lever, you get an output, you either keep it or pull again. Retake is the first tool that lets you stop pulling the lever and start directing — to look at footage and say “this is right except for this moment” and actually fix only that moment.
That’s the difference between generative art and generative production. The first is a happy accident. The second is a craft.
LTX-2 Retake makes the second possible.
Break it. Fine-tune it. Ship with it.
Have you tried LTX-2 Retake? Building something with it? Got a wild use case in mind? Let me know on X — drop the most ambitious thing you’d retake.