LTX-2.5 Requirements: VRAM, GPU & File Sizes (2026)

LTX-2.5 is an open-weight video-plus-audio generation model released on August 11, 2026 by LTX, the studio spun out of Lightricks. It pairs a 22B-parameter diffusion transformer (DiT) with a custom Gemma 4 12B text encoder, and natively handles text-to-video, image-to-video, video-to-video, text-to-audio, and audio-to-video at up to 4K HDR. The short answer: expect 16GB to 80GB of VRAM depending on setup, with total downloaded file size ranging from roughly 35.9GB to 71.4GB depending on which quantized checkpoint you pick. Community reports put the practical floor around an RTX 4090-class 24GB card; LTX has not published an official VRAM requirements table. Below is the full breakdown — from a quick-reference spec table to troubleshooting.

None of these figures come from an official spec sheet — they are built from measured Hugging Face file sizes plus community-reported working configurations. Whether a given setup actually runs depends heavily on your CUDA/PyTorch versions and the resolution/frame-count you request.

LTX-2.5 makes two significant jumps over prior LTX-family models. First, it generates video and audio together — text-to-audio and audio-to-video run inside the same model, removing the need for a separate audio pass. Second, it natively targets 4K HDR output, with dedicated upscaler weights (latent_upscale_models) shipped alongside the core checkpoints.

The backbone is a 22B-parameter diffusion transformer (DiT). Text understanding runs through a custom Gemma 4 12B encoder with an added projection layer. Checkpoints come in two flavors: distilled (a fixed 8-step schedule with CFG=1, built for fast inference) and dev/full (a fine-tunable general-purpose checkpoint).

LTX's own benchmark reports an image-to-video, 10-second, 720p clip generated in about 6.8 seconds on 2x NVIDIA GB200 using the distilled checkpoint. Via LTX's managed API, the same job at 1080p takes 23.7 seconds. Local hardware won't hit those numbers, but they're a useful reference point.

You don't need every file. Pick one quantized variant of distilled or dev, one matching text encoder, and the base VAE set — that's enough to run.

- BF16: Highest fidelity. Use this for fine-tuning or when validating output quality. Heaviest on both disk and VRAM
- INT8+convrot (ComfyUI-specific): Quantization tuned for ComfyUI workflows, roughly halving BF16 size while limiting quality loss
- NVFP4 (Blackwell / ltx-kernels): Built for Blackwell-generation GPUs like the RTX 50 series. Smallest disk and VRAM footprint, but requires the dedicated ltx-kernels
- FP8-cast (low-VRAM fallback): Downcasts BF16 weights to FP8 at inference time in PyTorch. Combined with CPU offload, some users report running it under 16GB VRAM (at a significant speed cost)

LTX shipped official ComfyUI workflow templates on day one. The minimal working combination is:

- diffusion_models: ltx-2.5-22b-distilled-transformer-nvfp4.safetensors (18.72GB)
- text_encoders: gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors (15.37GB)
- vae: ltx-2.5-video-vae-bf16.safetensors (1.47GB)
- vae (if you also need audio): ltx-2.5-audio-vae-bf16.safetensors (0.36GB)
- Total disk: approximately 35.9GB

Switching to the INT8+convrot setup means swapping in ltx-2.5-22b-distilled-transformer-comfy-int8-convrot.safetensors (21.50GB) for the diffusion model, bringing the total to roughly 38.7GB. For a broader look at local inference tooling beyond ComfyUI, see our local LLM inference engine comparison.

- 16GB VRAM class (e.g. RTX 4060 Ti 16GB): The community-reported floor. Requires NVFP4 or FP8 downcast plus CPU offload. Speed and resolution are heavily constrained
- 24GB VRAM (RTX 4090): The comfortable target for INT8+convrot. Multiple reports of stable 720p-class generation
- 32GB VRAM (RTX 5090): Blackwell generation, where native NVFP4 optimization (ltx-kernels) kicks in — more headroom on resolution and frame count
- 80GB VRAM class (A100/H100): The realistic choice for BF16 full setups, fine-tuning, or longer/higher-resolution generation — aimed at development workloads

The distilled checkpoint runs a fixed 8-step schedule at CFG (classifier-free guidance) = 1. Where a typical diffusion model needs dozens of steps and CFG scale tuning, distillation compresses the inference process itself, so generation is fast — a good fit for production use and batch generation.

The dev/full checkpoint is the general-purpose, fine-tunable model — the base you would use for additional training on your own dataset or for building LoRAs. It targets research and customization work, and needs more steps and more VRAM than distilled.

LTX-2.5 enforces hard constraints on output dimensions. Frame count must satisfy num_frames % 8 == 1, so valid values follow the pattern 1, 9, 17 … up to 121 (steps of 8, plus 1). Width and height must both be multiples of 32. The officially validated example is 544x960 at 121 frames and 24fps.

Ignoring these constraints either throws an error or triggers automatic cropping/padding that produces unintended output, so it is worth checking your parameters first when building a workflow.

As of this writing, there is no official statement or published benchmark on Apple Silicon or CPU-only support for LTX-2.5. Given the model's scale — a 22B DiT plus a 12B text encoder — running on an MPS backend is not theoretically impossible, but whether it performs at a usable speed is untested. If you're considering a Mac setup, check community issue trackers and Discord for the latest reports before committing.

We cover MiniMax H3's full requirements in our MiniMax H3 requirements guide. If you need simultaneous audio generation, LTX-2.5 is the natural pick; if you're comparing purely on video generation, MiniMax H3 is worth weighing too. For text-generation local deployment comparisons, see our Qwen3.8-27B requirements guide.

LTX-2.5 is released under the LTX-2.x Community License. Companies and individuals with annual revenue under $10 million can use it commercially, including in production, at no cost. Most small and mid-sized businesses in Japan fall well under that threshold, so in practice this covers the vast majority of local use cases. Organizations above $10 million in annual revenue need a separate Commercial Use Agreement — worth confirming before you build on it.

- Running out of VRAM (OOM): Switch to NVFP4 or INT8+convrot first. If that still fails, try FP8 downcast plus CPU offload (expect a large speed hit)
- Load failures from CUDA/PyTorch mismatches: The recommended stack is Python >= 3.12, CUDA >= 12.7, and PyTorch ~= 2.7. NVFP4 in particular depends on ltx-kernels, which assume a Blackwell-generation GPU and a recent CUDA — older stacks may simply not load it
- Frame count or resolution errors: Confirm num_frames % 8 == 1 and that width/height are multiples of 32. Start from the validated 544x960, 121-frame, 24fps combination before tuning parameters
- Generation is unusually slow: Check whether CPU offload is unintentionally active, or whether you accidentally loaded the BF16 full setup instead of a quantized one