FalSoteDiffusionAPI

FalSoteDiffusionAPI runs Sote Diffusion, an anime-tuned flavor of Stable Cascade that lives only on fal.ai's servers. If you've never heard of it, you're in the majority - this is the pack's deepest cut. It's Stable Cascade underneath, fine-tuned toward an anime aesthetic, and this node is the only place most people will ever touch it. You get the two-stage Cascade controls plus a workflow that's clearly built for quality-tag prompting: the defaults are literally newest, extremely aesthetic, best quality, as the positive and very displeasing, worst quality, monochrome, realistic, oldest as the negative.

Those defaults tell you everything about the intended audience. This is the anime-quality-tags school of prompting - the same "best quality, absurdres" lineage that carried over from Danbooru-style models. If that's your workflow, the node hands you a ready-made starting point; if it isn't, replace the defaults with your own.

How it works

Under the hood it's the Stable Cascade two-stage dance again: a first_stage_steps prior generates a compressed latent, second_stage_steps decodes it into the final image. It submits to fal-ai/stable-cascade/sote-diffusion with the safety checker off. One quirk worth knowing: this node sets the environment variable ANIME_STYLE_API_KEY from your key file, which is what the Sote endpoint expects - so if you're poking at it outside ComfyUI, that's the auth secret it wants.

The two-stage defaults skew a little hotter than the plain Cascade node: 25/10 steps, guidance 8.0, decoder guidance 2.0. That's the anime-tuned sweet spot as shipped; fiddle from there.

The inputs that matter

• prompt / negative_prompt - both real and both meaningful (Cascade is a CFG model, so negatives work - unlike the Flux nodes in this pack).
• width / height - 256–2048 in 8px steps.
• first_stage_steps / second_stage_steps - 25/10 defaults.
• guidance_scale - 8.0 default, on the high side - typical for anime models that want crisp adherence.
• decoder_guidance_scale - 2.0 default here, vs 0.0 on plain Cascade.
• seed, api_key - standard.

Single IMAGE output.

Install

Same pack, same setup:

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 at fal.ai/dashboard/billing.

Troubleshooting

• Output looks like generic anime, not your prompt - this is a niche model with a narrow training distribution. Quality-tag prompting helps; verbose scene description helps less.
• "realistic" in the negative is doing real work - that's intentional. The Sote fine-tune fights toward anime style, and the stock negative is tuned to keep photos out. If you want realism, this isn't the node for it.
• Cost - still billed per image on fal. The model is obscure, not free.
• If it errors, it's probably the key - remember it wants ANIME_STYLE_API_KEY; a generic fal key file usually works through this node, but if you're scripting it yourself, that's the var name to set.

The pack is archived. This node is a thin wrapper around a fal endpoint that could vanish anytime - Sote Diffusion has essentially zero community footprint, so enjoy it while it's up.