Replicate stability-ai/stable-diffusion

This is the classic Stable Diffusion - the original text-to-image model that made local image generation a real thing back in 2022 and launched the entire hobby this whole ecosystem grew out of. This node calls Stability's hosted version through Replicate, so you can fire off a generation with zero setup: REPLICATE_API_TOKEN, pay-per-run, nothing downloaded.

Let's be honest about where it sits in 2026: it's ancient. Four architecture generations have come and gone. Anatomy is shaky, hands are the meme, resolution is small, and anything modern will beat it on quality. So why is it here? Nostalgia, that unmistakable early-SD look, reproducing an old workflow, or just poking at the model that started everything. For serious work you'd reach for something newer - but as a piece of living history you can run in one node, it's fun.

How it works

Classic SD is a latent-diffusion model with a single CLIP text encoder - small, fast, and famously unpredictable. That unpredictability is part of the charm: sparse, noisy training captions left it with odd associations that read as "creativity," something later, better-captioned models trained out. The node just relays your prompt and settings to the hosted endpoint on Replicate.

The inputs and outputs that matter

• prompt - keep it under the 77-token CLIP limit; older SD can't hold complex multi-part scenes.
• negative_prompt - genuinely useful here (classic SD is a full-CFG model), and the standard way people fought its artifacts.
• width / height (both default 768, chosen from fixed sizes) - stay near native; push much larger and it duplicates subjects and breaks composition.
• num_inference_steps (50) and guidance_scale (7.5) - the classic SD defaults, CFG 7-ish being exactly right for this generation (unlike the low-CFG modern models).
• scheduler (DPMSolverMultistep default, plus DDIM, K_EULER, K_EULER_ANCESTRAL, PNDM, KLMS) - the sampler. DPMSolverMultistep is a fine default; K_EULER_ANCESTRAL is worth a try for variety.

num_outputs (up to 4) and seed round it out; force_rerun forces a fresh run. Output is one IMAGE.

How to install it

One install covers every Replicate node.

• ComfyUI Manager: search ComfyUI-Replicate, install, restart.
• Manual:

  cd ComfyUI/custom_nodes
  git clone https://github.com/replicate/comfyui-replicate
  cd comfyui-replicate
  pip install -r requirements.txt
  

  then restart.

Nothing lands on your GPU - the model runs on Replicate's servers, so the only install is the replicate Python client. Set your token before launching ComfyUI: export REPLICATE_API_TOKEN="r8_..." (replicate.com/account/api-tokens). No token, no runs.

Common issues & troubleshooting

Auth error first run. REPLICATE_API_TOKEN must be set before ComfyUI launches. Set it, restart.

Bad hands, deformed anatomy. That's classic SD, not you. It's a fundamental limit of the 2022 model; negative prompts help a little, nothing fixes it. If anatomy matters, use a modern model.

Composition breaks at high resolution. Stay near the native size and upscale afterward. Ask for a huge canvas directly and it clones subjects.

Same image on re-run. Replicate returned a cached result for identical inputs. Change the seed or set force_rerun.

Why pay for a model with free open weights? You wouldn't, if you already run SD locally - this node is pure convenience, no download, no VRAM. Each run is still a billed Replicate prediction.