⚡ Gigachad VAE Decode

Gigachad VAE Decode is the node at the end of every pipeline: it turns a LATENT back into an IMAGE. The stock VAEDecode does the same thing, but this version adds two things you'll actually hit: tiled decoding for big images, and automatic video-frame flattening so video latents come out as usable image batches.

The inputs that matter

• samples (LATENT) and vae (VAE) - the obvious two.
• tiled (default off) - the switch that changes your life on big generations. When on, the node uses vae.decode_tiled instead of a full decode, which trades a little quality for dramatically lower peak VRAM. On an 8GB card trying to decode a 2048px+ image, tiled is frequently the difference between "works" and "CUDA out of memory."
• tile_size (default 512) and tile_overlap (default 64) - the spatial tile size and how much tiles overlap to hide seams. If you see visible tile boundaries in tiled output, bump the overlap; if you're still OOMing, shrink the tile size.
• temporal_size (default 64) and temporal_overlap (default 8) - the same idea applied to the time axis, for video VAEs. Defaults are a sensible start for LTX/Wan-class video latents.

The single output is image (IMAGE), ready for a save or preview node.

What it does under the hood

Full decode is just vae.decode(). Tiled mode calls decode_tiled with your tile/overlap settings for both space and time. The part that's easy to miss: if the decoded tensor comes out 5-dimensional (B T H W C - a video latent), the node flattens it to B*T H W C automatically. That means your video frames come out as one big image batch instead of a shape that downstream image nodes choke on. That flattening is the difference between this node "just working" with video models and a stock VAE decode throwing shape errors.

Where it fits

Standard text-to-image: wire the KSampler's latent here and it's the final hop before Save Image. For video: it's the decode stage that hands frame batches to whatever saves frames. For hi-res work: it's where you flip tiled on. The KB's upscaling essay is worth a read in this context - decode is also where you start if you're re-encoding for img2img, and where you end when upscaling.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/comfyui-gigachad.git

or install comfyui-gigachad via ComfyUI Manager and restart. No requirements.txt in the pack - pure ComfyUI.

Honest verdict: it's a stock node with a tiled switch and a video convenience bolted on. If you only do 1024px SDXL stills, the stock VAEDecode is fine and this is interchangeable. If you generate big images on a tight card, or you're doing video at all, the tiled + flatten behavior is worth having. Same for its sibling - the GigachadVAEEncode in this pack mirrors these controls on the encode side.