EasyCascadeKsampler (Full)

Same honest framing as the simple version of this node: Stable Cascade had a real moment in early 2024, got squeezed between the SD3 announcement and Flux's launch, and the ecosystem mostly moved on before Cascade had time to accumulate the tooling investment its early benchmarks arguably earned. This is the "full" variant - where easy cascadeKSampler leans on settings already baked into your pipe, this one exposes steps, cfg, sampler, scheduler, denoise, and seed directly on the node itself, plus the encode/decode VAE handling Cascade's architecture actually needs.

Why Cascade needs its own VAE handling

Stable Cascade's three-stage Würstchen pipeline doesn't share one VAE the way SD and SDXL do - you're working with separate encode and decode paths as part of moving between the aggressively compressed stage-C latent space and something usable further down the pipeline. That's why this node, unlike almost anything else in the pack, has both encode_vae_name and decode_vae_name as required fields rather than a single vae_name. Get these swapped or mismatched and you'll get garbage output that looks like nothing else in your workflow is actually broken - it's specifically a Cascade quirk, not a general VAE mistake.

The inputs and outputs that matter

pipe carries your loaded Cascade model and conditioning in. The sampling fields - steps, cfg (4 default, notably lower than the SD-era 7-8 norm), sampler_name, scheduler, denoise, seed - work as you'd expect from any KSampler-equivalent, just exposed directly rather than pulled from a presampling config node. image_output, link_id, and save_prefix follow the same convention as every other Easy-Use sampler node - control what happens to the result (hide, preview, save, or route through the sender/receiver system) without a separate node bolted on.

Three optional inputs matter for stage-C specifically: image_to_latent_c lets you feed a source image to be encoded into the stage-C latent space rather than generating from scratch, latent_c takes a stage-C latent directly, and model_c overrides the stage-C model from what's in the pipe.

Three outputs: pipe, model_b, and latent_b. That last pair is the actual point of the "full" version - model_b and latent_b are what you hand off to a stage-B pass, since Cascade's decompression toward pixels happens in a separate stage from the structural stage-C sampling this node performs. This node is flagged as an output node, so it can preview or save directly depending on image_output, but the real output that matters for a complete Cascade pipeline is the model_b/latent_b handoff.

Installing it

ComfyUI Manager: search ComfyUI Easy Use, install, restart. By hand:

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

Windows: double-click install.bat. Linux/Mac: run pip install -r requirements.txt yourself. Restart ComfyUI. Cascade's stage B and stage C checkpoints, plus its encode/decode VAEs, are all separate downloads - budget for several gigabytes before you get a working pipeline together.

Where people get stuck

The VAE mismatch mentioned above is the single most Cascade-specific failure mode - if output looks like noise or static rather than a bad-but-recognizable image, check encode_vae_name and decode_vae_name before anything else. Second: forgetting that model_b/latent_b need to actually go somewhere - this node only runs the stage-C pass; if you stop here, you have a structural latent, not a finished image, and you need a stage-B sampling step downstream to decompress it. And given how little fresh Cascade content exists in the community at this point, if you hit an error that doesn't match anything above, it's worth checking whether it's a documented Cascade quirk from 2024 before assuming it's a new bug - most of what could go wrong with this architecture was already found and discussed while people were still actively building on it.