Sloyd: Sketch to Image

This is the node for people who can't draw. You scribble a chair - or take a rough line drawing, a blueprint, a whiteboard photo, a kid's dinosaur - and get back a finished, lit, textured image of that thing. Default prompt is literally "a wooden chair, product photo", which tells you exactly where the endpoint is aimed: concepts in, presentable product imagery out.

Why it's interesting next to the pack's other 2D nodes

Sloyd: Image Edit changes an existing image. Sketch to Image invents the material, lighting, and detail from a shape. That difference shows up in the plumbing too: the sketch goes up as a direct multipart file upload, not as an uploaded image job, which is why this endpoint is the one that takes multipart on Sloyd's side. You don't need to care - the node handles it - but it's why the tooltip says "First frame only" the same way the other nodes do.

The practical use inside this pack: sketch something rough, generate it here, then feed the result into Image to 3D to get a mesh. That's a genuinely short path from "I hacked out a shape on paper" to a GLB in Blender.

Inputs

• sketch - an IMAGE. Contrast matters far more than quality here: a high-contrast line drawing or a dark-on-white photo gives the model clean edges to work from, where a soft pencil sketch on grey paper is mostly noise. Thirty seconds in a levels node before this one is the best investment you'll make.
• prompt - up to 4096 characters. Name the object and the finish: "a wooden chair, product photo", "a brass desk lamp, studio lighting, white background". The prompt decides the material and the setting; the sketch decides the silhouette.
• seed - read the tooltip once and then never worry about it again. Sloyd has no seed parameter, so this doesn't reproduce anything - it forces a fresh generation by changing ComfyUI's input hash, and each fresh generation costs credits. Leave it alone and re-queueing an unchanged graph replays from cache for free.
• timeout_seconds - 900 by default. The poll runs every two seconds, shows a progress bar, and checks for interruption so Cancel actually stops it.

Outputs

image is a standard ComfyUI IMAGE - Save Image, previews, upscalers, all the usual consumers. image_path is the saved file relative to ComfyUI/output (e.g. sloyd/<job_id>.png) for anything outside ComfyUI. job_id is the raw Sloyd identifier. sloyd_job is the typed handle that chains into the rest of the pack, credentials included, so you never copy an id by hand.

Install and credentials

ComfyUI Manager → search Sloyd, or:

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

Restart ComfyUI. One Python dependency (requests), no model downloads, no GPU requirement - the sketch is uploaded and rendered on Sloyd's side. Then Settings → Sloyd → Client ID and Client Secret from api-dashboard.sloyd.ai. Two values, not one; billing must be enabled first (prepaid, $12 minimum, 75 credits per dollar); the secret is displayed once and cannot be retrieved later. For Docker or a headless runner, set SLOYD_CLIENT_ID and SLOYD_CLIENT_SECRET instead - that path is also the safe one if your ComfyUI listens beyond localhost, since the credential route has no auth of its own.

Things that bite

Your sketch is normalised on the way out: anything longer than 2048px on its long edge is downscaled, and a PNG over 20MB gets re-encoded as JPEG. That's plenty for a line drawing and not much for a detail-dense technical illustration - if the output is losing fine structure, the input was probably resampled before the model saw it. Only the first frame of a batch is transmitted, so a batch of four sketches is one sketch billed four times. And remember the meter: a 402 error names the shortfall and starts nothing, while every successful run is charged when the job begins, refunded automatically if it fails.