LTX ยท Diffusion transformer
LTX-Video
LTX-Video is a pipeline of 3 networks totalling 7.104 billion parameters, of which 1.923 billion do the actual generating. At its native 768x512 and 121 frames the denoiser works on 6,144 latent tokens at once, and attention over that sequence is what sets the render time.
Built for speed rather than fidelity, and the only video model that generates faster than it plays on a strong card.
What the pipeline is made of
This is the part that separates a media model from a language model: these are separate networks, and whether they sit in memory together is your choice, not the model's.
| Component | Parameters | Published as | On disk | What it does |
|---|---|---|---|---|
| transformer | 1.923B | F32 | 7.17 GB | Does the generating. Runs once per step, so it dominates both memory and time. |
| text_encoder | 4.762B | F32 | 17.74 GB | Turns your prompt into conditioning. Runs once, then can leave the GPU entirely. |
| vae | 0.419B | F32 | 1.56 GB | Converts between pixels and the compressed latent space. Small, but its decode pass is a memory spike. |
These figures were measured from the weight file headers.
Memory by precision
Everything resident, at the model's native output size, on a card large enough that precision is the only constraint.
| Precision | Peak VRAM | Quality | What it costs you |
|---|---|---|---|
| BF16 | 15 GB | 100% | How the weights are published. No loss, and the largest footprint. |
| FP8 | 8.78 GB | 97% | Halves the denoiser with a small, usually invisible cost. Needs Ada or newer. |
| GGUF Q8_0 | 9.17 GB | 98% | Works on any card, unlike FP8. Slightly slower than native precision. |
| GGUF Q5_K_M | 6.98 GB | 95% | A middle step when Q8 will not fit. |
| GGUF Q4_K_M | 6.31 GB | 91% | The usual way a 12B image model gets onto an 8 GB card. Detail softens. |
| NF4 | 5.97 GB | 89% | Aggressive 4-bit. Fast to load, noticeably looser on fine detail. |
Memory by arrangement
The other lever: the same weights at the same precision, moved around differently. On this model the strongest lever is neither: run the text encoder on the processor. It is 67% of the pipeline, and once it is off the GPU the 27% denoiser is all that remains.
| Arrangement | Peak VRAM | Time | How it works |
|---|---|---|---|
| Everything resident | 15 GB | 19 s | All components stay on the GPU. Fastest, and needs the most memory. |
| Text encoder on CPU | 5.64 GB | 20 s | The prompt is encoded once on the processor, so the encoder never touches the GPU at all. Standard practice on video models, whose encoders are often larger than their denoisers. |
| Component offload | 10.14 GB | 21 s | One component on the GPU at a time. The text encoder runs, then makes way for the denoiser. Costs a few seconds per generation. |
| Sequential offload | 2.77 GB | 1 min 41 s | Weights stream layer by layer from system RAM. Runs almost anything on almost anything, and is many times slower. |
What a bigger output costs
Both memory and time climb faster than the frame count, because attention is quadratic in the length of the latent sequence.
| Output | Latent tokens | Peak VRAM | Time |
|---|---|---|---|
| 480p, 3 seconds | 2,730 | 15 GB | 6.2 s |
| 480p, 5 seconds | 4,290 | 15 GB | 11 s |
| 720p, 3 seconds | 6,440 | 15 GB | 20 s |
| 720p, 5 seconds | 10,120 | 15 GB | 39 s |
| 720p, 10 seconds | 19,320 | 15.31 GB | 1 min 54 s |
Measured at BF16 with everything resident on an RTX 4090, so the columns compare with each other rather than with your machine.
Which hardware runs LTX-Video
118 of 118 consumer devices run it in some arrangement. Each row shows the fastest arrangement that fits on that device.
| Device | Memory | Verdict | Arrangement | Peak | Time |
|---|---|---|---|---|---|
| RTX PRO 6000 Blackwell NVIDIA | 96 GB | Comfortable | BF16 everything resident | 15 GB | 13 s |
| GeForce RTX 5090 NVIDIA | 32 GB | Comfortable | BF16 everything resident | 15 GB | 15 s |
| RTX 6000 Ada Generation NVIDIA | 48 GB | Comfortable | BF16 everything resident | 15 GB | 17 s |
| GeForce RTX 4090 NVIDIA | 24 GB | Comfortable | BF16 everything resident | 15 GB | 19 s |
| GeForce RTX 5090 Laptop NVIDIA | 24 GB | Comfortable | BF16 everything resident | 15 GB | 23 s |
| NVIDIA DGX Spark 128GB NVIDIA | 128 GB | Comfortable | BF16 everything resident | 15 GB | 24 s |
| GeForce RTX 5080 NVIDIA | 16 GB | Comfortable | BF16 everything resident | 15 GB | 26 s |
| GeForce RTX 4080 SUPER NVIDIA | 16 GB | Comfortable | BF16 everything resident | 15 GB | 28 s |
| GeForce RTX 4090 Laptop NVIDIA | 16 GB | Comfortable | BF16 everything resident | 15 GB | 29 s |
| GeForce RTX 4080 NVIDIA | 16 GB | Comfortable | BF16 everything resident | 15 GB | 30 s |
| GeForce RTX 5080 Laptop NVIDIA | 16 GB | Comfortable | BF16 everything resident | 15 GB | 31 s |
| GeForce RTX 5070 Ti NVIDIA | 16 GB | Comfortable | BF16 everything resident | 15 GB | 33 s |
| GeForce RTX 4070 Ti SUPER NVIDIA | 16 GB | Comfortable | BF16 everything resident | 15 GB | 33 s |
| GeForce RTX 4070 Ti NVIDIA | 12 GB | Comfortable | GGUF Q8_0 everything resident | 9.17 GB | 36 s |
| GeForce RTX 3090 Ti NVIDIA | 24 GB | Comfortable | BF16 everything resident | 15 GB | 36 s |
| RTX A6000 NVIDIA | 48 GB | Comfortable | BF16 everything resident | 15 GB | 37 s |
| GeForce RTX 4080 Laptop NVIDIA | 12 GB | Comfortable | GGUF Q8_0 everything resident | 9.17 GB | 39 s |
| GeForce RTX 3090 NVIDIA | 24 GB | Comfortable | BF16 everything resident | 15 GB | 40 s |
| GeForce RTX 4070 SUPER NVIDIA | 12 GB | Comfortable | GGUF Q8_0 everything resident | 9.17 GB | 41 s |
| GeForce RTX 3080 Ti NVIDIA | 12 GB | Comfortable | GGUF Q8_0 everything resident | 9.17 GB | 42 s |
| GeForce RTX 5070 NVIDIA | 12 GB | Comfortable | GGUF Q8_0 everything resident | 9.17 GB | 46 s |
| GeForce RTX 3080 12GB NVIDIA | 12 GB | Comfortable | GGUF Q8_0 everything resident | 9.17 GB | 48 s |
| GeForce RTX 3080 10GB NVIDIA | 10 GB | Comfortable | GGUF Q8_0 everything resident | 9.17 GB | 48 s |
| GeForce RTX 4070 NVIDIA | 12 GB | Comfortable | GGUF Q8_0 everything resident | 9.17 GB | 49 s |
| RTX A5000 NVIDIA | 24 GB | Comfortable | BF16 everything resident | 15 GB | 51 s |
| GeForce RTX 2080 Ti NVIDIA | 11 GB | Comfortable | GGUF Q8_0 everything resident | 9.17 GB | 52 s |
| GeForce RTX 5060 Ti 16GB NVIDIA | 16 GB | Comfortable | BF16 everything resident | 15 GB | 59 s |
| GeForce RTX 5060 Ti 8GB NVIDIA | 8 GB | Comfortable | GGUF Q5_K_M everything resident | 6.98 GB | 59 s |
| GeForce RTX 4070 Laptop NVIDIA | 8 GB | Workable | GGUF Q5_K_M everything resident | 6.98 GB | 1 min 1 s |
| GeForce RTX 4060 Ti 16GB NVIDIA | 16 GB | Workable | BF16 everything resident | 15 GB | 1 min 4 s |
| GeForce RTX 4060 Ti 8GB NVIDIA | 8 GB | Workable | GGUF Q5_K_M everything resident | 6.98 GB | 1 min 4 s |
| GeForce RTX 3070 Ti NVIDIA | 8 GB | Workable | GGUF Q5_K_M everything resident | 6.98 GB | 1 min 5 s |
| Jetson AGX Orin 32GB NVIDIA | 32 GB | Workable | BF16 everything resident | 15 GB | 1 min 6 s |
| Jetson AGX Orin 64GB NVIDIA | 64 GB | Workable | BF16 everything resident | 15 GB | 1 min 6 s |
| Radeon RX 7900 XTX AMD | 24 GB | Workable | BF16 everything resident | 15 GB | 1 min 6 s |
| Radeon PRO W7900 AMD | 48 GB | Workable | BF16 everything resident | 15 GB | 1 min 7 s |
| GeForce RTX 3070 NVIDIA | 8 GB | Workable | GGUF Q5_K_M everything resident | 6.98 GB | 1 min 9 s |
| GeForce RTX 5060 NVIDIA | 8 GB | Workable | GGUF Q5_K_M everything resident | 6.98 GB | 1 min 13 s |
| RTX A4000 NVIDIA | 16 GB | Workable | BF16 everything resident | 15 GB | 1 min 13 s |
| Radeon RX 7900 XT AMD | 20 GB | Workable | BF16 everything resident | 15 GB | 1 min 19 s |
| Radeon RX 9070 XT AMD | 16 GB | Workable | BF16 everything resident | 15 GB | 1 min 24 s |
| GeForce RTX 4060 Laptop NVIDIA | 8 GB | Workable | GGUF Q5_K_M everything resident | 6.98 GB | 1 min 24 s |
| GeForce RTX 3060 Ti NVIDIA | 8 GB | Workable | GGUF Q5_K_M everything resident | 6.98 GB | 1 min 26 s |
| Radeon RX 7900 GRE AMD | 16 GB | Workable | BF16 everything resident | 15 GB | 1 min 28 s |
| GeForce RTX 4060 NVIDIA | 8 GB | Workable | GGUF Q5_K_M everything resident | 6.98 GB | 1 min 31 s |
Source: Lightricks/LTX-Video . Downloaded 0.5 million times in the last month. See how we calculate, or browse every video model.