Stable Diffusion 1.5 (custom model)

Stable Diffusion 1.5 is the model that refuses to die - four years on it still has the largest checkpoint library in the ecosystem, and its 10,000+ community fine-tunes cover every niche from anime to architectural rendering. This node is how you use that entire library without downloading any of it: it's an architecture node, which means the model input takes the AIR of any SD 1.5 checkpoint in Runware's catalog and runs it in the cloud. The default even points at a legend - civitai:4384@128713, which is DreamShaper 8.

That "custom model" framing is the whole point of the Runware/Custom models category: one node per architecture (SD 1.5, SDXL, FLUX, and so on), model swapped at runtime via the search button. You don't install checkpoints; you rent them per generation.

The inputs that matter

• model (required, STRING, default civitai:4384@128713): the AIR of any SD 1.5 checkpoint. Use the search button to browse Runware's catalog - that's the part beginners miss, because the box looks like it wants a URL and actually wants a model ID.
• positivePrompt and negativePrompt (multiline): the usual pair. Worth knowing SD 1.5's CLIP encoder has a hard 77-token limit per prompt, so keep it tight - this isn't a T5 model.
• width / height (128–1024, step 8): SD 1.5 is 512-native, so the 1024 cap is generous and results up there get soft unless you're doing hires work. Keep the default 1024 only if you're using advanced_json for a hires fix - the tooltip says that's what it's for.
• steps, CFGScale, strength, maskMargin, clipSkip: all gated behind enable toggles. Flip one on to reveal its value. That's the pack's way of leaving the model's defaults alone unless you opt in.
• scheduler (76 options), promptWeighting (compel / sdEmbeds), vae: the SD1.5-era knobs, all present and correct. promptWeighting: compel gives you A1111-style (word:1.2) syntax.
• Feature sockets: seedImage, maskImage (img2img/inpaint), plus outpaint, lora, controlNet, ipAdapters, embeddings, ultralytics, watermark, acceleratorOptions - each takes a Runware builder's typed output. Stack them freely.
• numberResults: up to 20 - the biggest of any node in the pack, because SD 1.5 is cheap. This is the node to lean on for batch variation.

Output is a native image (IMAGE).

How it works

Same pack mechanism, one twist: the model widget overrides the default on every run, and the builders you wire in get stripped down to exactly the fields this architecture accepts before the request goes out. Your prompt, dims and any features are packed into a REST request; the checkpoint is loaded on Runware's side; the image comes back and lands as an IMAGE tensor in your graph.

Install and gotchas

Pack-wide install: ComfyUI Manager → search Runware, or git clone https://github.com/Runware/ComfyUI-Runware into custom_nodes + pip install -r requirements.txt, restart, add a key (Settings → Runware API key, RUNWARE_API_KEY, or runware auth login).

Where people get burned: pasting a CivitAI page URL into model instead of an AIR (use the search button), running prompts past 77 tokens and watching them silently truncate, and treating 1024×1024 as free when SD 1.5's sweet spot is 512. And because it's a paid API, numberResults: 20 is a deliberate spending choice - fun, but deliberate. For a huge, battle-tested checkpoint ecosystem with zero disk footprint, this is the node.