⭐ Star Latent Resize

Here's a failure mode you'll hit within your first week of img2img: you encode an image, wire the latent into a sampler, and get a cryptic shape mismatch - or worse, a silently wrong-looking result, because your latent's size isn't what the model's training expects. Star LatentResize fixes the boring half of that problem: it resizes a LATENT to a target resolution with one dropdown, and hands you back the exact new WIDTH and HEIGHT as numbers so nothing downstream has to guess.

How it works

Latents live at a fraction of pixel resolution (a 1024px image is a 128×128 latent after an 8x-downsampling VAE), but you don't need to think in that scale here. The node takes your latent and either preserves its aspect ratio or lets you specify a custom size, then interpolates to the chosen target. The ratio dropdown is the aspect decision: Keep Input Ratio (the default) scales to a target area while preserving your source's proportions, and Custom Size uses custom_width / custom_height directly. That distinction is the whole usability win - most resize nodes force you to pick a bare resolution and stretch your content to fit; this one can shrink-to-fit without squashing.

The resolution dropdown (21 options) is where the presets live, and they're model-aware rather than abstract pixels: SD (~512), SDXL (~1024), Qwen Image (~1328), WAN HD (1280×720), WAN FullHD (1920×1080), then 2–7 Megapixel targets. Pick the family you're sampling with and the node handles the divisibility. custom_width / custom_height default to 1920×1080 with a 16-pixel step, and the range goes up to 99968 - plenty for outpainting-adjacent workflows.

Inputs and outputs

• LATENT - your encoded latent.
• ratio - keep aspect or go custom.
• resolution - the model-aware size preset, or custom to use the width/height fields.
• custom_width / custom_height - used when ratio is Custom Size or resolution is custom.
• LATENT (out) - the resized latent, same channel count, ready for a sampler.
• WIDTH / HEIGHT (out, INT) - the actual new latent dimensions. Wire these into an Empty Latent or a conditioning node that needs to know the shape, so you never hardcode a number.

Where it fits

Anywhere a latent crosses model boundaries: img2img where your source is 832×1216 and the model's happy zone is 1024², or a latent upscale where you resize before a high-res pass. The WIDTH/HEIGHT outputs are the sneaky good part - they let downstream nodes adapt automatically instead of you re-typing dimensions.

Installing

Part of the StarNodes pack - install Starnodes via ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/Starnodes2024/ComfyUI_StarNodes
cd ComfyUI_StarNodes
pip install -r requirements.txt

Restart, then search star on the canvas. No models, pure torch interpolation.

Common issues

The trap is forgetting that resizing a latent is not free detail - enlarge too far and the sampler invents mush. Keep your target within the model family's native range (the presets exist for exactly that reason). And if a sampler still complains about shape after resizing, check you're not resizing to a size that breaks the model's divisibility rule; the 16-step on the custom fields is there to prevent it, so use the fields rather than free-typing odd numbers.