What it does

ComfyUI-PrunaVAED loads PrunaAI/PrunaVAED from ComfyUI's standard models/vae directory and exposes it as a normal ComfyUI VAE.

It replaces the LTX-2.3 video VAE in a workflow. It does not replace the LTX audio VAE or the LTX transformer.

LTX-2.3 latent ──> Load PrunaVAED ──> VAE Decode ──> video frames
                         │
                         └── standard ComfyUI VAE output

• Native ComfyUI model and VRAM offloading path
• Standard VAE output
• Compatible with regular VAE Decode
• Compatible with LTX tiled decode nodes that call the standard VAE interface
• Video encoding support for image-to-video workflows
• No Diffusers runtime dependency
• No automatic downloads, telemetry, or background network requests

Important

PrunaVAED is a video VAE only. Continue loading LTX23_audio_vae_bf16.safetensors with the regular LTX audio VAE loader.

Nodes

Load PrunaVAED

Loads and validates the PrunaVAED checkpoint.

Port	Direction	Type	Description
vae_name	input	filename	Selects a checkpoint from ComfyUI/models/vae
device	input	choice	main_device runs on the active accelerator; cpu runs entirely on CPU
weight_dtype	input	choice	Selects bf16, fp16, or fp32 for VAE weights and compute
VAE	output	VAE	Standard ComfyUI VAE object for video encode and decode

Internally, the loader remaps the upstream Diffusers-style state dictionary to the native ComfyUI LTX VAE layout and instantiates the pruned decoder channel structure. Unsupported checkpoints fail with an explicit format error.

main_device with bf16 is the recommended configuration and matches the upstream PrunaVAED usage. Choose cpu only when freeing accelerator VRAM is more important than decode speed. fp32 offers the widest CPU compatibility but uses twice the model memory of bf16/fp16. fp16 support and performance depend on the selected device and PyTorch build.

Benchmark LTX VAE vs PrunaVAED

Measures both decoders against the same latent tensor.

Port	Direction	Type	Description
samples	input	LATENT	One shared LTX-2.3 latent
reference_vae	input	VAE	Stock LTX-2.3 video VAE
pruna_vae	input	VAE	VAE returned by Load PrunaVAED
warmup_runs	input	integer	Untimed warm-up passes
timed_runs	input	integer	Passes used for the median
report	output	string	Per-run timings, medians, and decode speedup

CUDA is synchronized around every timed pass. Model loading is excluded after warm-up, and the final report is also written to the ComfyUI console.

Installation

1. Install the custom node

Clone or copy this repository into:

ComfyUI/custom_nodes/ComfyUI-PrunaVAED

No additional Python packages are required.

2. Install the model

Download the official checkpoint from the PrunaVAED model repository:

vae/diffusion_pytorch_model.safetensors

Rename it and place it in the standard ComfyUI VAE directory:

ComfyUI/models/vae/PrunaVAED.safetensors

3. Restart ComfyUI

Add Load PrunaVAED, select PrunaVAED.safetensors, and connect its output wherever the workflow expects the LTX-2.3 video VAE.

Leave device on main_device and weight_dtype on bf16 for the recommended GPU or accelerator configuration. Select cpu to keep VAE execution off the accelerator; use fp32 if the CPU or PyTorch build does not execute reduced-precision operators efficiently.

Workflow wiring

Text-to-video

LTX sampler output ──> VAE Decode
                           ▲
Load PrunaVAED ────────────┘

Image-to-video

Use the same VAE output for both source-image encoding and final video decoding.

Audio

Keep the existing audio path unchanged:

LTX23_audio_vae_bf16.safetensors ──> LTX Audio VAE Decode

Reproducible decode benchmark

Drag examples/PrunaVAED_benchmark.json into ComfyUI.

The workflow:

1. Creates one synthetic LTX-2.3 latent at 1280 × 768, 121 frames.
2. Sends that exact latent to the stock and Pruna decoders.
3. Runs warm-up passes.
4. Measures synchronized decode passes.
5. Prints median latency and calculated speedup.

This isolates VAE decode performance. It does not benchmark sampling, text encoding, audio generation, file encoding, or total workflow time.

Performance expectations

The upstream PrunaVAED model card reports approximately 1.7× faster video decoding and about 50% lower peak VRAM in its tested configuration.

Those numbers were measured upstream on an NVIDIA H100 using BF16 and mostly full-frame decoding without tiling or rolling. Results can differ substantially with:

• GPU architecture
• resolution and frame count
• ComfyUI's internal chunking
• spatial or temporal tiling
• dtype and VRAM mode
• model offloading and concurrent models

Only the latent-to-video decode stage is accelerated. If sampling dominates the workflow, the total generation time may change only slightly.

Local benchmark results

The benchmark node was also validated on an NVIDIA GeForce RTX 4090 using BF16, one warm-up pass, three synchronized timed passes, and median reporting.

Latent	Decoded video	Stock LTX-2.3 VAE	PrunaVAED	Speedup
1 × 128 × 16 × 24 × 40	121 × 768 × 1280	3,373.9 ms	1,898.8 ms	1.78×
1 × 128 × 51 × 24 × 40	401 × 768 × 1280	12,149.0 ms	6,177.9 ms	1.97×

This reduced measured decode latency by approximately 43.7% for 121 frames and 49.1% for 401 frames.

Raw synchronized timings

121 frames
Stock LTX-2.3 VAE: 3373.9, 3431.8, 3286.8 ms
PrunaVAED:          1843.4, 1899.6, 1898.8 ms

401 frames
Stock LTX-2.3 VAE: 12149.0, 12209.6, 12025.3 ms
PrunaVAED:          6132.3, 6177.9, 6267.5 ms

Compatibility

Component	Status
LTX-2.3 video latents	Supported
Standard ComfyUI VAE Decode	Supported
LTX tiled decode nodes	Supported through the standard VAE interface
Image-to-video VAE encode	Supported
LTX audio VAE	Separate; not replaced
Other video model latent formats	Not supported
Diffusers installation	Not required

The loader is architecture-specific and intentionally rejects arbitrary VAE checkpoints.

Windows and Linux

The node contains no hard-coded drive letters, path separators, shell commands, or Windows-only APIs. Model discovery is handled by ComfyUI's folder_paths, while device placement and offloading use ComfyUI's model-management layer.

• Windows: validated with ComfyUI Portable and CUDA
• Linux: supported by the same platform-neutral ComfyUI code path
• NVIDIA / ROCm: synchronization is used when PyTorch exposes the device as cuda
• CPU / MPS: the benchmark skips CUDA synchronization and uses the active ComfyUI device

A native Linux GPU benchmark has not yet been run in this development environment, but no OS-specific runtime path exists in the node.

Sources and attribution

• PrunaAI/PrunaVAED — model weights, architecture description, upstream benchmark, and usage terms
• Comfy-Org/ComfyUI — host application, VAE interface, model management, and LTX VAE primitives
• Lightricks/LTX-2 — LTX-2 model architecture and license
• Lightricks/ComfyUI-LTXVideo — LTX-specific ComfyUI workflows and tiled decode integrations

License

The custom-node source code in this repository is distributed under the GNU General Public License v3.0.

The PrunaVAED model weights are not included and remain subject to the LTX-2 Community License Agreement. Review the upstream model and license terms before use or redistribution.