⚡ Gigachad VAE Encode

Gigachad VAE Encode is the mirror image of the pack's VAE Decode: it takes an IMAGE and compresses it into a LATENT, which is what you need the moment your workflow stops being pure text-to-image. Load a reference photo, encode it, and feed the latent into a sampler at denoise < 1 - that's img2img, style transfer from an input, inpainting setups, and a dozen other patterns, all starting with this node.

The inputs

• pixels (IMAGE) and vae (VAE) - the image to encode and the VAE to do it with. Both required.
• tiled (default off) - tiled encoding, same logic as the decode side. Big source images (2048px+, or high-res photos you're feeding in for img2img) can OOM a small card on a full encode; flipping this on splits the work into tiles.
• tile_size (default 512) and tile_overlap (default 64) - spatial tiling controls. Shrink the tile size if you're still OOMing; raise the overlap if you see seams.
• temporal_size (default 64) and temporal_overlap (default 8) - temporal tiling for video sources, where you're encoding a clip rather than a still.

The single output is latent (LATENT), which wires into a sampler's latent_image or into a latent-manipulation node.

How it works

Full encode is vae.encode(); tiled mode uses the VAE's tiled encoder with your spatial and temporal tile settings. There's no magic here - it's the standard VAE compression with explicit memory controls exposed. The two things to actually remember:

1. Encode + denoise is the img2img recipe. Encode your source, drop denoise to something like 0.3–0.6, and the sampler preserves the structure while changing the style or refining detail. Full denoise = 1.0 mostly ignores the input.
2. Tiled encode is not free. It's a memory safety net, not a quality boost - keep tiled off for normal-size images and only flip it when you hit an OOM. The KB's concepts doc covers the same "tiled when you must" pattern for upscaling.

Install

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

or search comfyui-gigachad in ComfyUI Manager and restart. No requirements.txt - pure ComfyUI, nothing extra to install.

Verdict: a solid, slightly more capable version of the stock VAEEncode, worth having if you do any image-input work on modest VRAM. If your workflow is strictly text-to-image and you never start from a photo, you won't touch it - but the day you want to img2img a reference, this is the node that makes the latent. Pair it with the pack's GigachadVAEDecode and the controls match on both sides.