Stable Cascade Checkpoint Loader (Inspire)

Let's be honest up front: this is a node for a model that mostly didn't stick. Stable Cascade (the Würstchen-architecture release from early 2024) arrived with real promise and then faded fast as the community's attention moved to SDXL finetunes and then Flux. In our reddit corpus its mentions trickle down to essentially zero by mid-2026. So if you landed here wondering whether you need this node for a modern workflow - you almost certainly don't. It's here for the people still running Cascade, and for that it's genuinely convenient.

What makes Stable Cascade fiddly to load is that it's a two-stage model: a Stage C that does the heavy generation in a highly compressed latent space, and a Stage B that decodes that up. Loading both by hand is a chore. This node loads them together in one shot, and optionally caches them in Inspire's backend so you're not reloading on every run.

How it works

You point it at your stage_b and stage_c checkpoint files and it loads the whole Cascade stack - the two models plus their VAEs, the CLIP, and the CLIP vision needed for Stage C - and exposes them all as outputs. The optional cache_mode ties into Inspire's Backend Cache: pick none, stage_b, stage_c, or all to keep the loaded stages in memory so switching workflows doesn't reload multi-gigabyte checkpoints from disk. The key_opt_b / key_opt_c fields are the cache keys for that, mirroring how Inspire's Shared Checkpoint Loader works.

The inputs and outputs

The ones you set:

• stage_b and stage_c - dropdowns for the two checkpoint files. Both are required; Cascade needs both stages.
• cache_mode - none by default; set it to all if you're bouncing between Cascade workflows and want the models kept warm in the backend cache.
• key_opt_b / key_opt_c - optional cache keys, only relevant when caching.

The outputs are the full stack, named for what they are: b_model, b_vae, c_model, c_vae, c_clip_vision, clip, plus key_b and key_c (the resolved cache keys, usable later to remove the cached data). You wire these into a Stable Cascade sampling graph.

How to install it

ComfyUI Manager: Install Custom Nodes → search ComfyUI Inspire Pack → install → restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Inspire-Pack

Then restart. The node ships in the pack (Dr.Lt.Data's) with no extra dependencies - but you do need the Stable Cascade checkpoints themselves, the Stage B and Stage C files, placed in your checkpoints folder. Those are the multi-gigabyte downloads, not the node.

Common issues

Empty dropdowns mean the checkpoint files aren't found - Cascade's Stage B and Stage C models have to be in your ComfyUI checkpoints directory or the pickers stay blank. Second, don't cross the stages: Stage B goes in the stage_b slot and Stage C in stage_c, and swapping them will produce garbage rather than an obvious error. Third, if you enable caching, remember it's session memory that eats RAM/VRAM for as long as it's held - fine when you're actively switching Cascade workflows, wasteful if you set all and forget it. And the honest catch-all: if the model is giving you trouble, weigh whether Stable Cascade is even the right choice in 2026 before you sink time into it - a current SDXL or Flux workflow is where the community and the tooling actually are.