Empty Sana Latent Image

Same job as ComfyUI's core EmptyLatentImage, adapted for Sana's compressed latent space - this is the txt2img starting point you plug into your KSampler before anything gets denoised. Sana doesn't use SD's familiar /8-compression VAE, so it needs its own version of this node with matching assumptions baked in, and this is the Sana-specific one (as opposed to Empty DCAE Latent Image, this pack's other option, which lives under the generic latent category rather than tucked under Sana specifically).

The pack's README lists both nodes together for Sana txt2img without drawing a hard line on which pairs with which of Sana's two possible VAE downloads - reasonable, since the underlying compression math is closely related between them either way. If your output comes out visibly broken rather than just imperfect, a latent/VAE mismatch is the first thing to suspect, and swapping which of the two empty-latent nodes you're using is a cheap thing to try.

Inputs. Three, all required, with tooltips written by the node's own author. width - "The width of the latent images in pixels" - default 512, range 16 to 16384, step 8. height, identical tooltip and range, for the vertical dimension. batch_size - "The number of latent images in the batch" - default 1, range 1 to 4096.

Output. LATENT - "The empty latent image batch," per its own tooltip. Feeds straight into your KSampler.

Where it sits in a full Sana chain. Sana Checkpoint Loader gives you the model, Gemma Loader + Gemma Text Encode give you conditioning, this node gives you the starting latent, and ExtraVAELoader (loaded with one of Sana's special VAE files, not a standard SD VAE) decodes the result at the end. All four pieces have to agree with each other - get one wrong and the failure shows up as bad pixels, not a helpful error message.

Installing the pack. ComfyUI Manager: search "Extra Models for ComfyUI." Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI_ExtraModels

then pip install -r requirements.txt in your venv, restart ComfyUI.

Read this before you spend an hour debugging Sana. The pack's own README carries a caution box specifically on the Sana section: a meaningful number of people have run into problems, and the maintainer's stated advice is to try the fork built by the Sana team themselves, which auto-downloads the model files and is apparently more reliable end to end. A full integration rewrite was reportedly in the works but not ready at time of writing. This node isn't where that trouble tends to live - it's just a shape generator - but if you're building a Sana workflow from scratch and things aren't behaving, that's the first documented thing to check before assuming your graph is wired wrong. If you've already got PixArt or DiT running through this same pack and don't want to juggle two node sets, staying here and working through the caution notes is the more practical path, just go in knowing it's the rougher of the two options.