Empty DCAE Latent Image

If you've used ComfyUI's core EmptyLatentImage before, this is that same idea for a different, more aggressively compressed latent space - and using the wrong one of the two is a real way to get broken output that looks like something else is wrong.

DC-AE (Deep Compression AutoEncoder) is one of the VAE architectures Sana can run against, and it compresses much harder than a standard SD VAE - 32 channels at /32 compression, per the pack's own README, versus the SD VAE's more familiar /8. A latent tensor shaped for standard SD compression fed into a DC-AE pipeline (or vice versa) simply doesn't line up, and you get garbage rather than a clean error. This node exists purely to hand you a correctly-shaped empty latent from the start, so that mismatch never happens in the first place.

Notice this one lives under ComfyUI's generic latent category rather than tucked under ExtraModels/Sana - that's a small but real hint about scope. It's shaped for the DC-AE compression ratio specifically, not hardcoded to Sana, so it'll sit in your node search alongside the standard latent nodes rather than buried in a model-specific submenu.

Inputs. Three, all required, and all with tooltips straight from the node author. width - "The width of the latent images in pixels" - defaults to 512, ranges 16 to 16384, steps by 8. height is the same, same range, same step. batch_size - "The number of latent images in the batch" - defaults to 1, ranges 1 to 4096.

Output. LATENT, described plainly as "The empty latent image batch." Wire it straight into your KSampler's latent input, right alongside your Sana model and Gemma conditioning.

Which latent node pairs with which VAE. The pack's README lists this node and Empty Sana Latent Image together as the two txt2img starting-latent options for Sana, without spelling out a strict one-to-one pairing to the two available Sana VAE downloads. The safe read: if you're loading the mit-han-lab dc-ae-f32c32-sana-1.0 VAE through ExtraVAELoader, this node's compression assumptions match that architecture by name. If your latent and VAE genuinely don't correspond, you'll know quickly - decoded output will be visibly broken, not subtly off.

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 and restart.

A heads-up before you build a whole Sana workflow around this. The pack's own README puts a caution notice on Sana as a whole: enough people have hit trouble that the maintainer recommends trying the fork maintained by the Sana team instead, which handles model downloads more smoothly. This latent node itself is simple and unlikely to be your problem - but if the rest of your Sana pipeline is fighting you, that fork is the documented first thing to try before you assume something in your own graph is wrong.