FalStableCascadeAPI

FalStableCascadeAPI runs Stability's Stable Cascade through fal.ai and gives you back an IMAGE. Stable Cascade was Stability's attempt at an efficient text-to-image architecture - the "Wurstchen" design where a small stage generates a compressed latent and a second stage decodes it, so you get higher quality per parameter than the old UNet path. It never really caught on: it arrived in early 2024, got eaten by the SD3 disaster and then Flux, and by 2026 it's mostly remembered as "the one with the weird two-stage sampling." This node is a way to poke at it without committing to the ~16GB of files local inference wants.

How it works

The interesting thing about Cascade is the two-stage dance, and this node exposes it directly. There's a first_stage_steps for the prior stage (which does the actual semantic generation) and a second_stage_steps for the decoder stage (which upscales that latent into a full image). Because it's a classifier-free-guidance model, it also takes a real negative_prompt - something the Flux nodes in this pack can't do, since Flux's distilled tiers have no traditional CFG. If you've been missing negative prompts, this is where they live.

Defaults: 20 first-stage steps, 10 second-stage, guidance 4.0, decoder guidance 0.0. Those are reasonable; the second-stage dial defaults to 0 because the decoder stage generally doesn't want much guidance. The node submits to fal-ai/stable-cascade with the safety checker disabled and downloads one image.

The inputs that matter

• prompt / negative_prompt - the negative field is real and functional here, unlike Flux.
• width / height - 256–2048 in 8px steps. Cascade is an efficient architecture, so it tolerates higher resolutions better than the old 512-era models.
• first_stage_steps / second_stage_steps - the two-stage knobs above; 20/10 is a fine starting point.
• guidance_scale - 4.0 default, normal CFG behavior.
• decoder_guidance_scale - 0.0 default; rarely needs touching.
• seed, api_key - standard.

Single IMAGE output.

Install

The whole pack installs the same way:

cd ComfyUI/custom_nodes
git clone https://github.com/BetaDoggo/ComfyUI-Cloud-APIs

or Manager → search "ComfyUI-Cloud-APIs", restart, fal key in keys/, billing topped up at fal.ai/dashboard/billing.

Troubleshooting

• First-stage results look abstract / blobby - that's normal, the prior stage's latent isn't a picture yet. If the final image is blobby too, bump first_stage_steps a bit.
• Negative prompt seems to do nothing - at low second_stage_steps the decoder barely responds to conditioning. Raise second-stage steps if you're fighting artifacts.
• Nobody uses this model anymore - true, but that's also the appeal: it's cheap, it runs in the cloud, and it gives you the two-stage controls that the current default models don't expose. Treat it as an experiment, not a daily driver.

The pack is archived and unmaintained, but this is a thin API call - it keeps working until fal retires the endpoint.