🦊 RS Image MultiLatent

ComfyUI's stock Empty Latent Image works fine until it doesn't - its 8x downscale assumption is baked in, and the moment your VAE divides by 16, 32, or a Qwen's 28, you're fighting "size not divisible" errors or silently off-res latents. 🦊 RS Image MultiLatent (RS_Image_MultiLatent) is the replacement that reads the VAE's actual latent_channels and scale_factor and sizes the latent correctly, full stop. Wire in any VAE and it adapts, no configuration.

How it works

When you run it, the node inspects your connected VAE - via first_stage_config or the VAE's own latent_channels attribute - and picks up the true downscale factor (4, 8, 16, whatever it is). It also matches the VAE's device and dtype, which avoids those annoying half-cpu/half-cuda tensor errors when your VAE ended up on a different device than the sampler. Then it builds the latent in one of three ways:

• Preset - pick from a library of 100+ resolutions. There are three preset families (Standard, Qwen, Krea2) because those models like different grids; Qwen presets start at 1008px rather than 1024, which is the whole story of why stock latents misbehave on it.
• Custom - manual width/height from 8 to 4096.
• Megapixels - set a target like 1.5 MP plus an aspect_ratio, and it computes the resolution for you. This is the VRAM-control mode.

Inputs that matter

• vae - plug in any VAE. The whole point is that it doesn't assume 8x.
• size_mode - Preset, Custom, or Megapixels.
• active_preset - Standard, Qwen, or Krea2, which decides which preset_* dropdown is live.
• batch_size - 1 to 64, how many identical latents to stack.

Outputs: latent (the LATENT tensor for your sampler), plus width_px and height_px integers you can wire into a text overlay or a resize node if you want the rest of the workflow to know the dimensions.

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 by searching "ComfyUI_RaykoStudio". No extra Python packages beyond what the pack already pulls in (torch, numpy, Pillow, and friends).

Common issues

• Latent shape surprises. If you hand a 1024x1024 preset to a model that wants a 28-multiple grid (Qwen), pick the Qwen preset family instead - that's exactly why it's there.
• Batch output. batch_size repeats the same latent, which is right for "generate N from one seed slot" workflows but not for video-style frame batches.

The honest sizing-up: for plain SD 1.5 and SDXL, the stock Empty Latent node is fine and this is a luxury. It earns its keep the moment you're on Flux, Krea2, or any Qwen-Image architecture, where the default 8x math is quietly wrong. If you're only ever on one architecture, you can probably skip it.