StableCascade_EmptyLatentImage

Stable Cascade isn't one model - it's a two-stage pipeline. A 1B-parameter "prior" (Stage C) generates a tiny, heavily-compressed image, then a second, smaller decoder (Stage B) turns that into a full-size image. That split is why this node looks odd at first: it outputs two latents, stage_c and stage_b, and it's the one that produces the blank starting canvas for both stages. Most empty-latent nodes give you one tensor; Cascade needs a pair because the two stages live in different latent spaces with different compression.

How it works

The math is the whole story. stage_c is [batch, 16, height // compression, width // compression] - at the default compression of 42, a 1024×1024 image becomes a latent only about 24 pixels per side. That's the point: the prior plans the image in a tiny space, which is why Cascade was famously cheap to run. stage_b is [batch, 4, height // 4, width // 4] - a much less compressed latent that the Stage B decoder will eventually turn back into pixels. Both come out as zeros, ready for the sampler to fill in.

The inputs that matter

• width / height (INT, 256–16384, step 8, default 1024) - the final image size. These drive both latents' dimensions.
• compression (INT, 4–128, default 42) - how hard the prior's latent is squashed. Leave it at 42 unless you know the model expects otherwise; the Cascade checkpoints were trained at this value.
• batch_size (INT, 1–4096, default 1) - how many images in the run.

Outputs wire in like this: stage_c goes to the Stage C KSampler, whose output then feeds StableCascade_StageB_Conditioning and the Stage B sampler, which starts from stage_b. If you're doing image-to-image instead, you'd use StableCascade_StageC_VAEEncode rather than this node's blank canvas.

How you get it

Ships with ComfyUI core - nothing to install. It's been there since Cascade launched in early 2024, so any modern ComfyUI has it. You still need the actual models (the Stage C and Stage B checkpoints) loaded through the standard checkpoint/UNet loaders, plus the Cascade VAE.

Honest notes

Stable Cascade aged quickly - SD3 and Flux moved past it, and the community's main surviving memory of the architecture is that its second decoder could hallucinate details (people famously watched it re-interpret eye colors) rather than faithfully reconstruct. That makes it more of a curiosity in 2026 than a daily driver. But the two-stage structure was genuinely influential, and this node is the cleanest illustration of it: two latents, two samplers, one image. If you're poking at Cascade to understand how stage-based generation works, this is where the tour starts.