EasyCascadeLoader

Stable Cascade is a pain to wire up by hand. Its Würstchen-based architecture runs in three stages - C, B, and A - each with its own model file, and a vanilla ComfyUI Cascade workflow is a small forest of loader and encode nodes. easy cascadeLoader collapses all of that into one box: point it at your stage models, type your prompt, pick a resolution, and it hands you a ready-to-sample pipe.

Worth being honest about what you're loading, though. Stable Cascade was a February 2024 Stability release that never found an audience - our corpus shows 664 threads in 2024 collapsing to 10 in the first half of 2026. Flux arrived that August and took over, and Stability's own follow-ups went to SD 3.5. This loader is well built and does exactly what it says; it just serves a model that most of the community has parked. Reach for it if you specifically want Cascade's look, not as a default.

How it works

The node loads all three stages plus CLIP, optionally applies a LoRA, sets up your empty latent at the chosen resolution and Cascade's compression factor, encodes your prompts, and bundles the lot into a PIPE_LINE. From there you feed it into easy preSamplingCascade and a Cascade sampler. It's the Easy-Use loader pattern - everything-in-one - applied to the fiddliest model architecture the pack supports.

The inputs that matter

• stage_c and stage_b - the two model files that actually matter. Stage C is the big diffusion model; stage B upscales its output. Both dropdowns read from your checkpoint/unet folders, and per the pack's own notes they can be the checkpoint-format Cascade files from Stability's comfyui_checkpoints release. stage_a is the VAE stage and usually stays on its baked default.
• resolution - a big preset list (31 options) plus a width x height (custom) mode that hands control to the empty_latent_width / empty_latent_height fields. 1024x1024 is the default and a sane start.
• compression (default 42) - Cascade's headline trick is its aggressive latent compression, and this is that dial. Leave it at 42 unless you know why you're changing it; higher trades detail for speed.
• positive / negative - your prompts, encoded right here so you don't need separate CLIP Text Encode nodes.
• lora_name plus its strengths, and an optional optional_lora_stack input, if you're stacking LoRAs.

Outputs: pipe (feed the sampler chain), model_c, latent_c, and vae - broken out separately in case you want to wire the raw pieces into non-Easy-Use nodes.

How to install it

Via ComfyUI Manager, search "ComfyUI Easy Use" and install. Or:

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

Then install.bat / pip install -r requirements.txt, restart. Preinstalled on comfy.icu. The models are the real dependency - download the Stable Cascade stage-C and stage-B files from Stability's Hugging Face repo and drop them in your models folders first.

Common issues

Empty stage_c / stage_b dropdowns is the number one issue, and it's always the same cause: the model files aren't downloaded or aren't in a folder the node scans. Cascade doesn't ship with ComfyUI; you have to fetch it.

Second, keep the whole branch Cascade-native. This loader's pipe and latent are Cascade-specific - send them into easy preSamplingCascade and a Cascade sampler, not the regular ones. Mixing stages across architectures is the fastest way to a shape-mismatch error.