Img2Img Enhance + Upscale (Venice)

Upscaling usually means loading a local ESRGAN or SeedVR2 model, which eats VRAM you may not have. Img2Img Enhance + Upscale (Venice) does the job on Venice's servers instead: feed it an image, get a bigger - and optionally, generatively enhanced - image back. The trade-off is worth being clear about up front: this is a hosted, beta-ish endpoint from a one-person pack, so it's "wow that's convenient" one day and a 500 error the next. If you want it to just work forever, a local upscaler is still your friend.

The important distinction comes from how upscaling splits into two very different jobs: more pixels (pure interpolation - Lanczos, ESRGAN) versus more detail (a model inventing plausible content that wasn't there). This node does the second one, optionally. When enhance is on, Venice's image engine doesn't just resize - it re-renders detail, which means faces can change and textures get rewritten. That's the feature, not a flaw, but know what you're asking for.

The inputs

• image - the IMAGE tensor you want processed. The node takes the first image of a batch, so feed it a single image.
• scale - 1 to 4, default 2. This is where the first gotcha lives: the shipped code throws an error whenever scale is 1, regardless of the enhance setting, even though the tooltip promises scale 1 = "enhance only." As it stands, scale 1 is effectively unusable. Stick to 2–4.
• enhance - default false. Turn it on to let the enhancement engine modify the image during upscaling.
• enhance_creativity - 0 to 1, default 0.5. How much license the enhancer gets. 0 hugs your original; 1 is practically a new image.
• enhance_prompt - a short style phrase ("gold, marble", "angry, menacing"). The author's own advice: keep it short and descriptive. Capped at 1500 characters.
• replication - 0 to 1, default 0.1. How strongly the original's lines and noise are preserved. Higher values look less plastic but noisier.

Output

image - a standard IMAGE tensor, ready for a preview or Save Image node.

The gotchas, from the source

• The scale-1 path is broken as shipped (see above) - don't fight it, just use 2–4.
• Scale 4 on a large image will be dynamically dialed back by Venice to stay within size limits; you won't necessarily get a true 4x.
• The author's own globals.py carries a comment wondering whether the upscale endpoint even works yet. It does enough of the time to be useful, but treat failures as "the API being the API," not something you misconfigured.
• The node needs torchvision (for the tensor↔PIL conversion), which isn't in the pack's requirements.txt - normally fine, because ComfyUI ships it. If your install lacks it, pip install torchvision.

Installing it

Part of ComfyUI-Venice-API - install the pack once (ComfyUI Manager, search "ComfyUI-Venice-API", or clone into custom_nodes and pip install -r requirements.txt), restart, and put your venice.ai key in Settings > VeniceAI > API Key. No models to download; the whole thing runs remotely, which is exactly the point on a modest GPU.