🦊 RS Image to Latent

VAEEncode works, until it doesn't: your image is 1037x1091, the VAE wants multiples of 8, and you get a tensor that's been silently padded or an outright error. 🦊 RS Image to Latent (RS_ImageToLatent) encodes an image into latent space and fixes the size problem in the same step. It detects what the VAE actually needs, rounds your image to a valid size, and hands you the latent plus the exact pixel dimensions it used. For img2img and inpainting workflows, that removes a whole class of fiddly pre-processing.

How it works

The pipeline is: read the image → decide the target size → resize if needed (via comfy.utils.common_upscale) → vae.encode() → return. The size logic is the interesting part. divisibility_mode defaults to auto, which inspects the VAE's actual downscale factor instead of assuming 8 - so an SDXL VAE gets 8, a Flux VAE gets its real divisor, and so on. rounding_mode controls how the rounding happens: auto picks the nearest valid size, shrink only ever reduces (never upscales - good for keeping VRAM low), and expand only increases. mode then governs the whole target:

• auto - preserve the original aspect and size, just round to the required multiple.
• preset - 50+ common resolutions (square, portrait, landscape).
• custom - your own width/height (64–4096, step 8).
• megapixels - a target like 1.5 MP plus an aspect_ratio, useful for staying inside VRAM budgets.

Inputs that matter

• image, vae - the obvious two. Tooltip on vae: "VAE model for encoding (auto-detects divisibility)".
• mode - auto / preset / custom / megapixels.
• divisibility_mode - auto, or force 8/16/32/64/112/128.
• rounding_mode - auto / shrink / expand.
• upscale_method - lanczos (default), bilinear, nearest-exact, bicubic.
• batch_size - repeat the same latent N times.

Outputs: latent, plus width_px, height_px, width_latent (width/8), and height_latent (height/8) - the latter two are handy for downstream sizing math.

How to install

Part of ComfyUI_RaykoStudio:

cd ComfyUI/custom_nodes
git clone https://github.com/Raykosan/ComfyUI_RaykoStudio.git

Restart ComfyUI, or install via ComfyUI Manager (search "ComfyUI_RaykoStudio"). Uses only torch and the pack's standard deps.

Common issues

• Image came back slightly resized when I wanted it identical. That's rounding_mode = auto doing its job - your source simply wasn't a multiple of the VAE's divisor. Set it to shrink to guarantee no upscaling, or pre-crop with RSCropImage's multiple mode to keep control of the exact pixels.
• Megapixels mode output looks off-aspect. The aspect ratio list is fixed; if your composition doesn't fit one, use preset or custom instead.
• VAE encode/decode drift. Encoding is lossy - every encode/decode cycle costs a little sharpness (the KB's VAE section covers this). Don't chain encode→sample→decode→re-encode more than you need to.

The honest take: VAEEncode is fine if your images are already clean multiples. This node earns its keep the moment you're feeding odd-sized crops, phone screenshots, or anything from a batch selector into a model with strict requirements - it's the size-multiple problem solved at the right layer.