Ard 4Vae Decode

The VAE is what turns a latent back into pixels you can actually look at, and in the four-tile workflow you finish with four latents that all need decoding. Ard 4Vae Decode does that in a single node: feed it one VAE and up to four latents, and it returns four images. It's four VAE decodes wearing a trench coat, which is either exactly what you want or a waste of a node slot, depending on how much you value a tidy graph.

The inputs: samples_1 is required, samples_2 through samples_4 are optional - so it also works fine for one, two, or three latents, and unused outputs come back as None. vae is the VAE to decode with, and tile_size (default 512, step 64) is the interesting one.

That tile_size is why this node is worth more than its trivial wrapper suggests: it uses vae.decode_tiled, decoding the latent in tiles rather than all at once. Latent space is 8× smaller per side than pixels, so a 512 tile size means a 64×64 latent tile per decode pass. For a big image on a modest GPU, tiled decode is the difference between "it works" and an out-of-memory crash - the KB's troubleshooting material on VRAM covers exactly this pattern. Set tile_size roughly in line with what a single decode pass can fit; smaller tiles = less VRAM, and 512 is a safe starting point.

Outputs are image_1 through image_4, matching the input order. From there they typically feed Ard 4Img Combine for the final stitch, or go straight to a Save/Preview.

Gotchas worth knowing:

• The four latents don't need to be the same size - each is decoded independently with the same tile setting.
• If you only connect samples_1, the other three image outputs are None; wire downstream nodes accordingly or you'll get connection errors on the unused ports.
• This is decode-only. If you need to encode pixels back into latents, that's a different node.

Honest verdict: the built-in VAE Decode (Tiled) does everything this node does for a single latent, and four of them would do what this does in one shot - just with four nodes cluttering your canvas. If you've already bought into the pack's four-tile pipeline, this node slots in cleanly and gives you tiled decode as a bonus. If you decode one latent at a time, there's no reason to switch.

Install is the standard pack route - ComfyUI Manager → search ComfyUI-Ardenius, or:

cd ComfyUI/custom_nodes
git clone https://github.com/ArdeniusAI/ComfyUI-Ardenius

then restart ComfyUI. Pack-wide notes: civitai and moviepy come along in requirements, and a stray import error at startup is the Save Image node wanting comfyui_controlnet_aux - the rest of the pack still loads.