Stable Diffusion Latent Upscaler

There's an old distinction in the upscaling world that still matters: more pixels versus more detail. The first is what ESRGAN and Lanczos do - they make the image bigger without inventing anything. The second is what generative upscalers do - they run the image through a model that adds plausible detail where the upscaled pixels are empty. This node (runware:502@1) is the classic Stable Diffusion latent upscaler, the second-pass workhorse from the SD 1.5 era, running on Runware's cloud. You feed it an image, it comes back 2x bigger, with real added texture instead of just interpolated mush.

If that sounds modest next to the FLUX nodes in this pack, it is - and that's sort of the point. The latent upscaler is a utility. It's the detail-add pass you chain after a generation, the "hires fix" intermediate, the thing that buys you a bigger canvas before a final refinement. It does one job, does it cheaply, and asks nothing of your GPU.

How it works

The mechanism is the SD latent upscale trick: the input image is encoded into latent space, the latent is scaled up (a factor of 2 doubles width and height), and then a diffusion pass re-renders it at the larger size, adding detail consistent with what's already there. Because it works in latent space with the SD 1.5 upscaler model, it's fast and light - but it's also a 2023-era model. It adds detail in SD 1.5's idiom, not FLUX's. Think of it as a dependable first rung on the upscale ladder, not a SUPIR replacement.

Inputs that matter

• image (required) - whatever you're scaling, as a plain IMAGE socket.
• upscaleFactor - the dropdown, default (default); the concrete choice is 2 (doubles width and height).
• settings.steps - toggle-gated, 20–60, default 20. More steps = more detail time.
• settings.CFGScale - toggle-gated, 3–12, default 3. This is the img2img pass's guidance; keep it low.
• settings.positivePrompt / negativePrompt - the detail pass has its own prompt (e.g. "detailed, sharp" style cues), separate from the original generation's.
• settings.clipSkip - 0–2, the classic CLIP-layer skip from the SD1.5 bag of tricks.
• settings.seed, outputFormat, outputQuality - the usual.

The image output is a native IMAGE, 2x the input, ready to feed the next stage or straight to Save.

The honest notes

• It adds SD 1.5-flavored detail. On a FLUX or Z-Image render, the latent upscaler's idea of "detail" can fight the source's style. It's great on SD-family images; on modern backbones you may prefer a modern generative upscaler.
• The (default) factor - if you leave it, you're trusting the model's default; the concrete choice in the dropdown is 2. There's no 4x here, so plan your upscale ladder accordingly (2x → 2x if you need 4x).
• Prompts actually matter on the pass. A positive prompt pointing at "detail" plus a negative that keeps artifacts out does noticeably better than empty strings.
• Toggle pattern, again. settings.steps, settings.CFGScale, settings.seed are all enable-to-set booleans with paired _value fields.

It's the least glamorous node in the Runware catalog and one of the most useful, because a working 2x latent detail pass is something every serious workflow eventually needs. Here it costs a fraction of a cent and zero VRAM.