🐸 Empty Latent

Frog Empty Latent replaces the stock EmptyLatentImage with a dropdown of resolutions somebody actually picked for you. The stock node makes you type width and height and remember that ComfyUI latents are one-eighth the pixel dimensions. This one gives you nine presets, all of them sensible for modern anime models like Anima and Flux, and hands you the width and height back as integers on top of the latent. That last bit is the part people don't expect: w and h come out as separate INT outputs, so conditioning or upscaler nodes that need to know the canvas size can read it straight off the wire instead of you hard-coding numbers in three places.

The presets are the argument for the node. ComfyUI is full of "empty latent" replacements now - it's one of the most-reinvented wheels in the ecosystem - but this list is noticeably not the SDXL-era set. Instead of 1024×1024 and a few landscape/portrait afterthoughts, you get 1536×1536, 1728×1344, 1344×1728, 1856×1248, 1248×1856, 2016×1152, 1152×2016, 2304×1024, and 1024×2304. Those are the resolutions that Anima actually likes - the KB notes it runs comfortably at 1536² even on a 6GB card - and each is a sensible aspect ratio (4:3, 3:4, 16:9, 9:16, 9:4). You pick, you generate, you don't do arithmetic.

The inputs

Only two, and only one you'll touch:

• preset - the resolution dropdown, defaulting to 1536 × 1536. The whole node is this.
• batch_size - how many empty latents to create in one tensor. Keep it at 1 unless you're deliberately generating a batch.

What comes out

latent (a zeroed latent at the chosen resolution), plus w and h as INTs. Wire latent into your sampler, and wire w/h anywhere that needs the canvas size - some conditioning and regional nodes consume it.

Installing it

Part of Frog Node Pack. ComfyUI Manager → search "Frog Node Pack", or clone into custom_nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/RabbitThatIsPink/FrogNodePack

Restart, hard-refresh, no extra dependencies - the pack runs on Pillow alone.

Where people trip

The only real trap is treating this as a universal resolution reference for SDXL models. These presets are sized for the 2B-generation flow-matching models - Anima and Flux - where native resolution runs higher than SDXL's 1024 sweet spot. Drop a 2304×1024 canvas onto an SDXL model and you'll get doubled artifacts; drop it onto Anima and it's a normal Tuesday. And if you're the type who wants a custom aspect ratio, this node doesn't have a free-form option - the presets are the feature. For that, the 🐸 LLM Latent Selector picks from this same list automatically based on your prompt, and 🐸 Smart Latent Switch lets the LLM choose with a manual fallback. This is the manual fallback.