🌫️ LDM Empty Latent

If you've ever built a txt2img workflow in ComfyUI, you know the Empty Latent node: it defines the blank canvas that the sampler will paint into. This is that node, sized for the CompVis LDM F8 Large instead of SD. It creates a zero-initialized latent tensor and hands it to the 🎲 LDM Sampler, which adds the noise and gets to work.

The reason it exists as its own node rather than being hidden inside the sampler is the same reason it exists in stock ComfyUI: the canvas dimensions are a decision you make before sampling, and they're baked into the shape of the tensor the whole run works with. Here they're baked in as a 4-channel tensor at 8× downsampling - a 256×256 canvas becomes a 32×32 latent - because the f8 VAE compresses each spatial dimension by a factor of 8.

The settings that matter

• width and height - both default to 256, and that default is not an accident. 256×256 is the resolution this model was trained at. It's the single most important knob in the whole pack, honestly: this node's job is to keep you honest about it.
• batch_size - how many images per pass, 1 to 8. Leave it at 1 while you're learning; each image is a full denoising run.

The 256 trap

Every article about this model says the same thing because it's the thing that actually bites people: the F8 Large was trained on tight, downscaled images, and it falls apart as you push the canvas up. At 512 you start to see duplicated structure and repeating artifacts; at 1024 it gets worse. That's the difference from SD1.x, which was fine-tuned to handle bigger canvases - this model never was.

The author's own guidance is blunt about it: constrain the canvas, e.g. 256×256. If you want a big image, generate at 256 and upscale afterward - that's the workflow that produces something clean, and it's what a modern upscaler is actually for.

Wiring and gotchas

Output is a single LATENT socket, feeding the sampler's latent_image input. Nothing else accepts it - the LDM_LATENT type is custom to this pack.

Two things worth knowing. First, the latent it hands the sampler is zeros, not noise - the sampler injects the noise itself, so don't expect a "grainy start" to show up in a preview. Second, if you must run img2img or try a bigger canvas, keep the dimensions a multiple of 8 or the VAE will complain about the tensor shape. The node's sliders step by 64, which always satisfies that, but hand-editing the value is where you can trip.

Keep it at 256, at batch 1, and let the rest of the chain do its thing. It's a two-minute node when you respect its one rule.