Draw Things Hints

In normal ComfyUI, using ControlNet means building a chain: load a checkpoint, run a depth or pose preprocessor node, wire the condition image into a ControlNet loader, then into the sampler. This node is the Draw Things answer to that whole chain, and it collapses most of it into one box.

DrawThingsHints bundles up to four images with a type and a weight, and hands the bundle to the Draw Things engine, which runs the preprocessing itself. You hand it a raw photo and say "this is a pose," and the server converts it into the condition its ControlNet actually needs. No ControlNet Auxiliary Preprocessors pack required - the preprocessing happens on the other side of the gRPC pipe.

One caveat up front, same as every node in this pack: the repo is deprecated in favor of the official draw-things-comfyui extension. The node still works; it's just frozen.

How it works

The node's add_hints function builds a hint stack. Each hint is a type plus a tensor plus a weight. When the sampler runs, it batches every hint of the same type together and sends them to the server as HintProto entries; Draw Things does the actual hint preprocessing (depth maps, pose skeletons, scribbles, color palettes, shuffle) and feeds them into the ControlNet or Kontext model. The tooltip in the node spells out the audience: hint images are used with ControlNets and certain other models, like Kontext.

The big win over a classic ComfyUI setup: you can feed it a raw image, not a preprocessed map. The trade is that you don't get to pick your depth-estimator or pose detector - that choice lives inside Draw Things.

The inputs that matter

The node gives you four slots, each with the same two widgets:

• type - (None selected), Depth, Pose, Scribble, Color, Shuffle (Moodboard), or Custom. This is the important one. Custom means "I already preprocessed this myself," which is the escape hatch when you want a specific detector.
• weight - 0 to 1, how strongly that hint pushes the generation. Start at 1 and dial down to loosen.

And four image inputs, one per slot. The trap: if you plug an image in but leave its type at (None selected), the node silently skips it. Set the type first.

There's also an optional hints input on the node, so you can chain hint bundles together. The DT_HINTS output wires straight into the sampler's hints input.

Installing it

ComfyUI Manager (search ComfyUI-DrawThings-gRPC) or:

cd ComfyUI/custom_nodes
git clone https://github.com/Jokimbe/ComfyUI-DrawThings-gRPC

Restart, and remember the server side needs to be configured: Draw Things app with API Server on, gRPC protocol, TLS enabled, Model Browser enabled, and Response Compression disabled (or run gRPCServerCLI with --no-response-compression --model-browser).

Troubleshooting

Hints doing nothing usually means one of two things. Either you left the type at (None selected) and the image got dropped, or the model you picked doesn't consume hints at all - a plain SD/SDXL checkpoint has no use for them, so hints only bite on ControlNet-backed and Kontext-style models. Also worth knowing: for a Shuffle (Moodboard) hint, Draw Things expects the reference collage, and shuffle models respond well to the target_blocks setting on the ControlNet node. If your hints feel too strong or too weak, tune the hint weight first, not the ControlNet weight - they multiply.