pruna_ltx2.3_vae_comfy_bf16.safetensors

pruna_ltx2.3_vae_comfy_bf16.safetensors is Kijai's ComfyUI-layout conversion of PrunaVAED, Pruna AI's accelerated decoder for the LTX 2.3 video VAE. The decoder channel structure is pruned; the encoder side and the latent space are identical to the stock VAE, which is why it is a drop-in swap — the same latents decode through either file.

Native support landed in ComfyUI core (commit c38171d, 'Support Pruna LTX VAE'), so on a current ComfyUI build the standard VAELoader node loads it directly. Older builds need the third-party ComfyUI-PrunaVAED custom node, which remaps the upstream Diffusers-style state dict to the native LTX VAE layout.

It is 1.33 GB BF16 — slightly smaller than the 1.45 GB stock BF16 video VAE — and lives in ComfyUI/models/vae/ alongside taeltx2_3 and the audio VAE. It replaces the video VAE only; audio pipelines still load LTX23_audio_vae_bf16.safetensors separately.

Choose the Pruna VAE when decode time or decode-step VRAM is the constraint: long or high-resolution clips where the final decode is a large share of total time, or a 16GB card that generates fine but OOMs when decoding. Pruna's headline numbers are ~1.7× faster decode and roughly half the peak decode VRAM against the stock BF16 VAE.

Stick with taeltx2_3.safetensors if you just want the smallest file that every published workflow already references — it is 23 MB and the default everywhere. Stick with LTX23_video_vae_bf16 if you want the unmodified full-fidelity decoder for HDR or archival output and don't care about decode speed.

All three coexist in models/vae/; switching is a dropdown change in the VAELoader node, not a workflow rebuild.

Some published workflow JSONs reference this file under a custom subdirectory. If ComfyUI shows a "cannot find model" error and your workflow references one of these path-prefixed variants:

• vae/pruna_ltx2.3_vae_comfy_bf16.safetensors
• ltx23\pruna_ltx2.3_vae_comfy_bf16.safetensors

The prefix before the slash or backslash is a subdirectory the workflow author used. The actual file is the same pruna_ltx2.3_vae_comfy_bf16.safetensors — you have two fixes:

1. Create the matching subdirectory inside ComfyUI/models/vae/ and place the file there. Example: if the workflow references vae/pruna_ltx2.3_vae_comfy_bf16.safetensors, create the corresponding subfolder under ComfyUI/models/vae/ and put pruna_ltx2.3_vae_comfy_bf16.safetensors inside it.
2. Or open the workflow JSON in a text editor and replace the prefixed string with just pruna_ltx2.3_vae_comfy_bf16.safetensors. ComfyUI then resolves it directly from ComfyUI/models/vae/.

On Windows the separator is \, on macOS/Linux it is / — they refer to the same nested folder regardless of platform.