PreSampling (Cascade)

This is the pre-sampler tuned for Stable Cascade - Stability's Würstchen-based model from early 2024 with the three-stage (C/B/A) architecture. It plays the same role every Easy-Use pre-sampler plays: it writes your sampling settings into the pipe without rendering. The Cascade-specific part is that it also handles the VAE names for encoding and decoding across Cascade's compressed latent stages, which a plain pre-sampler doesn't know about.

Level-set first, because it changes whether you should care: Stable Cascade never really took off. In our reddit corpus it pulled 664 threads across 2024 and just 10 in the first half of 2026 - a release spike followed by near-silence. Flux landed in August 2024 and ate the entire "big new open model" conversation, and Stability's own attention moved to SD 3.5. So easy preSamplingCascade is a node for a model most people have moved on from. If you're specifically running Cascade, it's a genuine convenience. If you're just browsing, you almost certainly want easy preSampling or a Flux/SDXL setup instead.

How it works

You pair this with easy cascadeLoader, which builds the pipe and hands you Cascade's stage-C latent. This node configures the stage-C sampling pass - the one that does the heavy lifting in Cascade's pipeline - and passes the pipe onward to a Cascade sampler (easy cascadeKSampler / easy fullCascadeKSampler). Because Cascade works in a heavily compressed latent space, the encode/decode VAE choices live on the node rather than being implied.

The inputs that matter

• steps (default 20) and cfg (default 4) - note the cfg default is 4, lower than the plain pre-sampler's 8. Cascade responds to lower guidance than SD 1.5/SDXL, so trust that default before cranking it.
• sampler_name (default euler_ancestral) and scheduler (default simple) - the Cascade-friendly defaults are already set. There are 44 samplers and 9 schedulers available.
• encode_vae_name and decode_vae_name - the stage-A VAE used to move in and out of pixel space. Usually left at the baked default from the loader.
• denoise (1.0 for txt2img) and seed - same as always.

Optional image_to_latent_c and latent_c let you seed the stage-C latent from an image or an existing latent for img2img. Output is the pipe.

How to install it

Install the pack via ComfyUI Manager (search "ComfyUI Easy Use") or clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Use

Then install.bat or pip install -r requirements.txt, and restart. Already installed on comfy.icu. Separately, you'll need the actual Cascade model files - stage_c and stage_b checkpoints from Stability's stable-cascade release - placed where the loader can find them, or this whole branch has nothing to run.

Common issues

The most common "it's broken" report is really "I don't have the models." Cascade ships as separate stage-C and stage-B files, and the loader's dropdowns come up empty if they aren't in your models folders. That's a download problem, not a node problem.

Beyond that, don't mix this with the standard easy preSampling - Cascade's latent stages aren't interchangeable with a normal SD latent, and feeding a Cascade pipe into a non-Cascade sampler (or vice versa) will error. Keep the Cascade branch pure: cascadeLoader → preSamplingCascade → a Cascade sampler.