Your LTX-2.5 Download Is Only a Third of the Bill

How to size LTX-2.5 to the machine you actually own, including the three files the download page does not mention and the gate that will stop you halfway.

There is a version of LTX-2.5 that fits in 12.6 GB. Somebody published it on Wednesday night, along with six other sizes, and the file names make it look like a simple choice: pick the number under your memory, click, generate video with sound on your own machine.

It is not a simple choice, because that 12.6 GB file cannot make anything by itself. It is one of four pieces, and one of the pieces it needs is bigger than it is.

This is the single most common way local video generation goes wrong, and it produces a specific kind of frustration: you measure your card carefully, you pick conservatively, you get an out-of-memory error anyway, and you conclude your hardware is not good enough. Usually your hardware is fine. You just paid for the wrong third of the model.

Here is how to size the whole thing properly, using the set of files that went up on September 16 as the worked example.

What LTX-2.5 actually is, and why it comes in parts

LTX-2.5 is Lightricks' open video model. It makes video with matching audio from text, from a still image, or from other video, and it can hold a character and a lighting setup across several connected shots in one pass. It is one of the two or three genuinely usable open video models right now, it runs on consumer hardware, and the licence is unusually kind, which I will come back to.

Like most video models, it is built in separate specialised pieces rather than one monolith:

1. The generator. The big one. It does the actual imagining.
2. The reader. Turns your prompt into something the generator understands. In LTX-2.5's case this is a Gemma 4 12B language model with a small extra layer bolted on that maps its output into the shape the generator expects.
3. The video decoder. Turns the generator's compact internal output back into actual pixels.
4. The audio decoder. Same job for the soundtrack.

Every one of those has to be in memory. The compressed downloads that people share are almost always only the generator, because that is the piece that compresses well and the piece that varies by hardware. The reader and the two decoders come from the original repository, at their original sizes.

That is the whole trap. Somebody posts "LTX-2.5 in 15 GB," you have a 16 GB card, and you think you are set. The reader alone is 15.4 GB in its smaller form.

The actual size table

comatto/LTX-2.5-Distilled-GGUF went up on September 16 at 22:51 UTC with seven compressed versions of the generator. These are the maintainer's own published sizes:

File	Size	Their note
LTX-2.5-Distilled-Q3_K_S.gguf	12.6 GB	Smallest, highest quality loss
LTX-2.5-Distilled-Q3_K_M.gguf	12.9 GB	For strict memory limits
LTX-2.5-Distilled-Q4_K_S.gguf	15.3 GB	Balances speed and limited memory
LTX-2.5-Distilled-Q4_K_M.gguf	15.7 GB	Recommended baseline
LTX-2.5-Distilled-Q5_K_M.gguf	18.1 GB	Strong prompt adherence
LTX-2.5-Distilled-Q6_K.gguf	18.6 GB	Very low quality loss
LTX-2.5-Distilled-Q8_0.gguf	23.6 GB	Nearly indistinguishable from the original

Now the three files that are not in that repository, and are not in that table. All of these live in Lightricks/LTX-2.5, and I confirmed every filename below is present in that repo. One caveat on the numbers: the sizes I quote for these come from the compressed upload's own card rather than from Lightricks, because Hugging Face does not expose file sizes through the interface I used. Treat them as good enough to plan with and not as gospel to the megabyte.

• The reader, in text_encoders/. Two options: gemma4-12b-with-proj-ltx-2.5-bf16.safetensors at 26.3 GB, or gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors at 15.4 GB. Take the second one unless you have a lot of room. This is the file people forget.
• The video decoder, vae/ltx-2.5-video-vae-bf16.safetensors, 1.47 GB. There is also a lighter, faster ltx-2.5-video-vae-conv-bf16.safetensors at 1.45 GB.
• The audio decoder, vae/ltx-2.5-audio-vae-bf16.safetensors, 365 MB. Skip this only if you genuinely do not want sound, which for a model whose selling point is synchronised audio would be a strange choice.

So the real arithmetic for the recommended setup is roughly 15.7 plus 15.4 plus 1.47 plus 0.365, which is about 33 GB of model files, not 15.7. If you are on 24 GB of card memory, you are going to be moving pieces in and out rather than holding them all at once, and you should plan for that rather than discover it.

Put this into practice

1. Accept the gate first, before you download anything. The upstream Lightricks repository is gated. You have to click through a consent screen on the Hugging Face page, which also signs you up for marketing you can unsubscribe from. Do this first, in a browser, while logged in. If you skip it and start scripting downloads, your reader and decoder fetches will fail with a permissions error while the ungated compressed generator sails through, which is a confusing five minutes.

2. Pick your generator by what you have left over, not by your total. Work backwards. Subtract the reader and the decoders from your memory budget, then choose the generator that fits the remainder. On a 24 GB card with the 15.4 GB reader loaded, Q3_K_S at 12.6 GB is not going to sit alongside it. You will be offloading. On 48 GB you can hold the recommended Q4_K_M plus everything else with room for the actual generation.

3. Get the workflow files from the same repository. This is the part I appreciate about this upload: it ships two ready ComfyUI setups, video_ltx2_5_t2v_GGUF.json for text-to-video and video_ltx2_5_i2v_GGUF.json for animating a still image. Load one of those instead of wiring loaders by hand. Hand-wiring a split-component video model in ComfyUI is an hour of your life and three wrong guesses about which loader node wants which file.

4. Add the upscalers only after you have a clip you like. Lightricks publish two optional second-stage files: ltx-2.5-latent-spatial-upscaler-x2-bf16-1.0.safetensors at 996 MB doubles your resolution, and ltx-2.5-latent-temporal-upscaler-x2-bf16-1.0.safetensors at 262 MB doubles the frame rate, so 24 frames per second becomes 48. Generate small and rough first, decide the shot works, then upscale. Running these during exploration just makes every wrong idea more expensive.

5. Check you are pulling from the repository you think you are. The README in this upload links to Abiray/LTX-2.5-Distilled-GGUF throughout, not to itself. It reads like a re-upload where the links were left alone. The files are all present in the repo I linked, but if you click through the README's own download links you will end up somewhere else. Worth knowing, not worth panicking about.

The licence is the good news, and it is specific

Most open video models this year have come with a licence that turns out to mean "open for learning, closed for earning." LTX-2.5 does not.

The LTX-2.x Community License permits commercial and production use at no cost for entities under $10 million in annual revenue. That threshold covers essentially every freelancer, small studio and two-person production company I know. You can put this in client work. You do not need to ask anyone.

That is a meaningfully different position from what a lot of the other open releases offer right now, and it deserves saying plainly rather than burying it in a footnote. It is also worth reading yourself rather than trusting my summary: the agreement is in the LTX-2 repository, and the two model cards I looked at point at slightly different paths for it, so go to the upstream one.

Where this breaks

Compression costs you something, and the maintainer says so. Their own notes describe Q3_K_M as having the highest quality loss and Q8_0 as nearly indistinguishable from the original. There is no published comparison grid showing you what that means on a real shot, so the honest answer is that nobody has quantified the middle of that table for you. If a job depends on it, generate the same prompt at two tiers and look.

Prompt style matters more than it should, and the model card admits it. Lightricks' own limitations list says prompt following is heavily influenced by prompting style, and that the model may produce artifacts in complex scenes. That is not a hedge, it is a warning. A prompt that works beautifully on one video model can return mush here. Budget time for learning this model's particular preferences, and read the official multishot prompting guide on the upstream card before you conclude the model is bad at something.

Memory offloading is slow, and "it fits" is not the same as "it is usable." If you are shuffling a 15.4 GB reader and a 15.7 GB generator through 24 GB of card memory, your generation time is going to be dominated by moving data rather than making pictures. It will work. It may not be a loop you want to iterate in. If your card is genuinely small, the smaller compressed tiers plus the lighter convolutional video decoder are the combination to try, and you should expect to be rendering rather than sketching.

None of this is a first-run experience. Four files from two repositories, one consent gate, one workflow import. Call it forty-five minutes if nothing goes wrong, plus however long your connection takes to move 33 GB. It is not hard. It is just not one click, and anybody who tells you it is has not done it recently.

The thing I would actually change about how this gets shared

Every compressed model upload I see publishes a size table for its own files and says almost nothing about the other pieces. Some mention them. This one mentions them, which is why I used it as the example. Almost none of them publish the number that matters, which is the total memory the whole assembly needs to sit in at once.

That number is knowable. It would take the person doing the compression ten minutes to measure and one line to publish. Its absence is why the most common complaint about local video generation is "my card is not good enough," when quite often the card was never the problem.

If you maintain one of these repositories, add the total. If you use them, ask for it. And if you have measured what the recommended tier actually needs in practice on a card you own, post the number somewhere, because right now the community is collectively guessing at a figure that only needs measuring once.

---

Medium metadata

• Title: Your LTX-2.5 Download Is Only a Third of the Bill
• Subtitle: How to size LTX-2.5 to the machine you actually own, including the three files the download page does not mention and the gate that will stop you halfway.
• Tags: AI Video, ComfyUI, Open Source, Video Production, Generative AI
• Suggested hero image: four unequal jigsaw pieces on a workbench, only one of them price-tagged, the largest one unlabelled and pushed slightly out of frame