LTX 2.5 VAE: The Diffusion Decoder, the Conv Fallback, and the Audio VAE

LTX-2.5 changed how video gets decoded, and it changed what "VAE" means in the LTX stack. Where 2.3 had a single hidden VAE inside a bundled checkpoint, 2.5 ships three separate official VAEs: a diffusion video decoder for quality, a convolutional decoder for speed, and an audio VAE with a built-in vocoder. Picking the wrong one costs you quality or speed, and skipping the audio VAE breaks audio entirely.

This guide explains the three files, what each costs, and how the choice is made in practice. Sourced from the LTX-2.5 model card, the LTX-2 GitHub README, and the ComfyUI LTX-2.5 tutorial. Checked August 13, 2026.

The three VAEs in LTX-2.5

The key insight: "the VAE" in 2.5 is not one thing, and the quality difference is not a rounding error. The official model card lists a new diffusion video decoder as one of its headline changes — it replaces the VAE reconstruction stage, with the stated effects of sharper faces, textures, and on-screen text, better motion, and fewer artifacts in demanding scenes.

The diffusion decoder: why it is the default

The GitHub README describes the diffusion decoder (NADiffusionDecoder) as the improved-quality option: better quality at the cost of longer decode time and more VRAM.

Two details matter in practice:

• It is the file the official workflows use. The ComfyUI T2V/I2V/FLF2V workflows all load ltx-2.5-video-vae-bf16.safetensors — the diffusion decoder, not the conv one. The ltx-pipelines Quick Start likewise uses the bf16 diffusion VAE.
• Speed depends on the natten backend. The README says the natten extra is the fastest backend for the diffusion video decoder — and it is Linux + CUDA only. On Windows and macOS it is skipped automatically and decoding falls back to a Triton or eager implementation; the same install command works everywhere (uv sync --extra natten), it just decodes slower on non-Linux.

So on Linux + CUDA with natten installed, the quality gap costs you the least. On Windows or macOS, the diffusion decoder is still the quality pick — just be aware the decode step will be slower than the conv fallback.

The conv decoder: the speed escape hatch

ltx-2.5-video-vae-conv-bf16.safetensors is the convolutional decoder: lighter, faster, no extra dependencies. The README frames it as the alternative when you want the decode step to cost less.

When it is the right call:

• Your decode step dominates your iteration loop (you are generating a lot of clips and decode time matters more than per-frame crispness)
• You are on Windows/macOS where the diffusion decoder loses its natten speedup anyway
• Your VRAM is tight and you need to shave every possible byte during decode

When it is the wrong call: anything where on-screen text, faces, or texture fidelity is the point of the render. The diffusion decoder is the official quality path, and the vendor's artifact claims (sharper faces, legible text, fewer smears in fast motion) are tied to it.

The audio VAE: not optional in an audio-video model

The third file, ltx-2.5-audio-vae-bf16.safetensors, is the audio VAE plus vocoder. It is a separate component with a separate job: encoding and decoding the audio branch that LTX-2.5 generates in sync with video.

Why people get burned: the file looks like an optional sibling of the video VAEs, but it is not. The official file lists include it in every generation set — ComfyUI's workflows place it in models/vae/ and load it, and the ltx-pipelines Quick Start passes it via --audio-vae-path. Skip it and you get video with no audio, or a pipeline error depending on the workflow.

If your output is silent B-roll and you genuinely never want audio, you can technically leave the audio VAE out of a custom minimal setup — but the official paths all include it, and dropping it is a custom configuration you own, not a documented mode.

Which one should you download?

The honest summary: start with the official default set (diffusion video VAE + audio VAE), and only switch the video VAE to the conv version if your iteration loop or VRAM demands it. The diffusion decoder is where the quality claims live.

Curious about the quality difference but not ready to download 66 GB? Generate with LTX 2.5 — audio included — in your browser at ltx23.app and judge the decoder quality on your own prompts.

FAQ

Why are there two video VAEs in LTX-2.5? The diffusion decoder (DiffVAE) is the new quality path — sharper faces, textures, on-screen text, fewer artifacts. The conv decoder is a lighter, faster fallback with no extra dependencies.

Which VAE do the official workflows use? The diffusion decoder (ltx-2.5-video-vae-bf16.safetensors) in both ComfyUI's official workflows and the ltx-pipelines Quick Start, plus the audio VAE.

Do I need the audio VAE? Yes, for any official workflow — it carries the audio branch (VAE + vocoder) that generates synchronized audio with video.

What is natten and do I need it? The neighborhood-attention backend that makes the diffusion video decoder fast. It is Linux + CUDA only; on Windows/macOS the decoder falls back to Triton or eager implementations (slower, but the same command works).

Can I use LTX 2.3's VAE with 2.5? No — the README is explicit that files are not interchangeable between the two models. 2.3 files don't carry over.

Bottom line

LTX-2.5's VAE layer is three files with three jobs: the diffusion video decoder (quality default, natten-fast on Linux), the conv decoder (speed/lightness fallback), and the audio VAE with vocoder (required for audio). Start with the official pair — diffusion video VAE + audio VAE — and swap the video VAE to conv only when decode speed or VRAM forces the trade. Test the decoder quality risk-free at ltx23.app before committing to the 66 GB local setup.

Sources

Verified August 13, 2026 against primary documentation:

1. Lightricks/LTX-2.5 — Hugging Face model card — VAE file list and notes (diffusion decoder, conv VAE, audio VAE + vocoder), new diffusion video decoder headline feature
2. Lightricks/LTX-2 — GitHub README — NADiffusionDecoder description, natten backend (Linux/CUDA only), conv fallback, audio VAE requirement, file interchange warning
3. ComfyUI official LTX-2.5 tutorial — VAE files in official workflows, models/vae/ placement
4. LTX-2 optimization documentation (GitHub) — diffusion VAE decoder backends

Decoder quality claims are vendor-described; perceptible differences vary by scene content and hardware. Verify on your own renders.