The films in this post
In short. Yes, you can make photoreal AI films on your own PC with LTX-2.5, an open-weights video model from Lightricks. We run it on one NVIDIA RTX 5070 Ti with 16 GB of VRAM. The PC costs about ₹1.5 to 1.6 lakh once, and the video model has no per-video or per-second fees. What still costs money every month is the Claude plan that directs the work and the image plan that draws our storyboards. The films above were all made this way.
What PC do you need to run LTX-2.5 offline?
This is the PC in our room above the shop in Vaniyambadi, as measured:
| Part | What we have |
|---|---|
| GPU | NVIDIA GeForce RTX 5070 Ti, 16 GB VRAM |
| CPU | AMD Ryzen 7 9700X, 8 cores and 16 threads |
| RAM | 32 GB (two 16 GB sticks) |
| System | Windows 11 Pro |
A PC like this costs about ₹1.5 to 1.6 lakh, once. That is Nav's figure at November 2026 prices, and shop prices change. We run the distilled 22B version of LTX-2.5 in 8-bit (int8) inside ComfyUI. Small scripts, written and run by Claude Code, drive it, so we never click through the graph by hand. There is more on the whole setup on AI films on your own PC.
What still costs money every month?
Not everything is offline. We are honest about the rest:
- A Claude plan. Claude Code is the director that writes and runs every script. Pro is about ₹1,800 a month, paid to Anthropic.
- An image plan. Storyboards and stills come from ChatGPT Images through Codex, on a ChatGPT plan. On the Plus plan we hit the image limit at 75 images in one usage window while making NETA on 14 November, and the reset came later than we had estimated.
- Electricity. The GPU peaked at 73 °C and 277 W in our render logs.
The video model itself runs on the PC. Nobody charges us per video.
How do we go from storyboard to video?
We call it the director's loop. For every film:
- Recall what worked on earlier films.
- Brief. Write what the film is and what rules it follows.
- Voice first. We record or generate the voice before any picture, because the voice sets the clock.
- One storyboard sheet per sequence. Codex draws a single sheet (3x2 panels for vertical, 2x2 for wide) and we cut it into panels. One sheet means every shot shares the same cast, light and grade.
- Animate each panel with LTX-2.5 in shots of 4 to 7 seconds.
- Over-generate, then trim before the point where the AI starts to drift.
- Upscale 2x with a Lightricks IC-LoRA, grade, add sound and music, and mix to -14 LUFS.
- Critic agent. A separate agent watches the cut and scores it before anything ships. We fix and review again. On one film, two critic rounds took the score from 5.5 to 7.5 out of 10.
How long does one LTX-2.5 shot take on an RTX 5070 Ti?
About a minute. Here is what we measured:
| Measure | Result |
|---|---|
| Benchmark, four named jobs, int8 | 61 s average cold, 66 s average warm |
| One 15-second portrait image-to-video at 0.6 megapixels | 111 s |
| 87 renders from Trishul episodes 3 to 5 and Kolar | 5,554 s in total, 63.8 s per render on average |
| Trishul episode 4 alone | 59 renders, 3,638 s (about 61 minutes of GPU time), for an 81.6-second final cut |
A typical render is one shot of about 5 seconds (121 frames at 24 fps). A 15-second shot uses 361 frames. Trishul is our five-episode Hindi sci-fi series and Kolar is a four-episode Tamil street drama. Their logs are the source of these numbers.
Do not plan on one render per shot. We make roughly three to four renders for every shot that stays in the final cut. Fifty-nine renders for an 82-second cut of shots about 5 seconds long is the example.
Which LTX-2.5 tools do we use, and which film shows each?
The films above play in full, so you can judge each tool yourself.
| Tool | What it does | Where to see it |
|---|---|---|
| Image-to-video | Animates one storyboard panel | Three Wonders, the first film above |
| Text-to-video | Makes a shot from words alone: the crowd seen from behind | NETA, the second film |
| First-and-last frame | Builds impossible shots between two images: the lake's eye in Three Wonders, the garage reveal in Basement, Please | Three Wonders and Basement, Please (the fourth film) |
| Hold (camera lock) | Uses the start image as a middle or end guide, which stops the camera wandering and fixes most AI drift | NETA |
| Slow motion | Stretches a moment | NETA |
| Layout-to-render | We block the camera path in Blender with grey boxes and LTX-2.5 renders it photoreal | Stashed: Out of Office, the fifth film |
| Audio-to-video | The recorded line drives a talking head | Our founder-show pilot (not on the site yet) |
For the talking head, record the line first. The audio drives the face. We used one take per speaker, two image holds at 0.34, and the Hindi pilot ran at 0.85 megapixels. Night Recovery, the last film above, is a 14.7-second shot made with the same pipeline.
Which prompt rules worked?
- Name the lens and say the tripod is locked.
- Ask for tiny physical motions with a visible cause, and let them settle before the cut.
- Let one prop carry the story.
- Spend the pixels on eyes and hands.
- Keep a long fixed sentence about light and geometry at the end of every prompt, so shots match.
- Cut on the spoken word.
- For a product, photograph the real product and use the real photo for the packshot. Add exact prices and logos in the edit. Never trust the model with text.
What goes wrong with LTX-2.5, and what is the fix?
| What fails | The fix |
|---|---|
| A face drifts away from the cast between shots | Cut the shot, or re-render it with hold |
| Objects morph (a stack of cash turned into something else) | Re-take with hold and a plain instruction |
| Hands melt | Re-take with hold |
| An animal lunges at the lens | Re-take with hold and a plain instruction such as "stays seated upright" |
| Text and logos garble | Add them in the edit |
| Music hits found by ear are wrong | Detect hits on the music file itself. In NETA a "hit" at 17.02 s was the quietest moment and the real hit was at 18.44 s |
| Frames drift by 3 to 4 when segments are cut by seconds | Count frames cumulatively |
The full story of the last two is in the NETA making-of.
Can you sell what you make with LTX-2.5?
Lightricks' community licence allows commercial use for companies under a revenue limit. The limit was US$10 million a year when we checked. Read the current licence before you sell work, because terms change.
Voices have their own rule. We use Nav's own voice, cloned offline (IndicF5 for Tamil and Hindi, Chatterbox for English), or stock voices from HeyGen and ElevenLabs. Clone only your own voice, or a voice you have written permission for.
FAQ
Is 16 GB of VRAM enough for LTX-2.5?
For us, yes. The distilled 22B model in 8-bit runs on a 16 GB RTX 5070 Ti. We have not tested other cards, so we cannot say what a smaller one does.
Can I run it on a laptop?
The films we draw in code, the 2D ones, run on an ordinary computer. AI video with LTX-2.5 needs an NVIDIA GPU. We have only measured this on a desktop RTX 5070 Ti.
Is this really offline?
The video model, the render and the edit run on the PC. Claude Code and the Codex storyboards need the internet and a paid plan.
How many films can one PC make?
In the eight days from 12 to 19 November 2026 we finished more than 60 films and cuts on this one PC. That count includes language cuts and versions. What we made, 12 to 19 November lists them.
Learn it
The 2BFT Lab teaches this: ₹10,000 once for the live course (live every Wednesday and Sunday, 7 PM IST, with every recording) or ₹1,000 for the recordings alone. The AI films session is on 7 December, the short-series session on 14 December and 3D on 28 December. You need your own Claude plan, which is about ₹1,800 a month for Pro, paid to Anthropic and not included. See the Lab, prices, the director skill and the films library.
