Hint Image Enchance JK🐉

The spelled-off-by-a-friend name is HintImageEnchance JK, and it exists because of a quiet disaster in ControlNet and video workflows: you preprocess a pose, depth, or canny map, then resize it to generation resolution with a naive resize, and the edges go mushy. For canny that's fatal - a smeared edge map is a worse control signal than no map at all. This node is a port of the high-quality hint-resize logic from ControlNet AUX (originally written by Lvmin Zhang), wrapped up as a JakeUpgrade node, and it's the difference between "good enough" and "actually clean" control maps.

How it works

You feed it a hint_image - your preprocessed control map - plus the target image_gen_width and image_gen_height. Then you pick a resize_mode:

• Just Resize - stretch to the target dimensions. Fast, but distorts aspect ratio.
• Crop and Resize - inner-fit: scale up so the whole target is covered, then crop the overflow.
• Resize and Fill - outer-fit: scale to fit inside the target, then letterbox with a border color sampled from the image edges.

The "high quality" part is the resize itself. The node detects whether the map is binary, a one-pixel-edge map, or a segmentation map and picks the interpolation accordingly, so thin canny lines and pose skeletons don't blur into 3px smudges when you blow them up 4×. It also handles 4-channel inpaint masks correctly (the alpha channel is treated as a mask, not blended like a color).

There are two extra outputs worth knowing: METADATA and MODE, both strings describing what was done - handy if you're building API workflows and want a record of the processing in your logs.

Inputs and outputs

• hint_image - the control map to enhance.
• image_gen_width, image_gen_height - the target size (defaults 512×512).
• resize_mode - Just Resize / Crop and Resize / Resize and Fill.
• IMAGE - the resized map.
• METADATA, MODE - descriptive strings.

Installing it

Ships in ComfyUI-JakeUpgrade, which pulls in opencv-python as a dependency - that's the heavy one this node rides on:

• ComfyUI Manager → search "JakeUpgrade" → install → restart.
• Or:

cd ComfyUI/custom_nodes
git clone https://github.com/jakechai/ComfyUI-JakeUpgrade
cd ComfyUI-JakeUpgrade
pip install -r requirements.txt

Windows portable: run install.bat or ../../../python_embeded/python.exe -s -m pip install -r requirements.txt. If the node errors on cv2, the opencv install didn't take - pip install opencv-python directly.

Common issues

• Edges still blurry. If you're feeding a pre-thinned edge map, "Just Resize" may still soften it - the high-quality path helps but doesn't perform miracles. Re-thin after resize, or preprocess at the target resolution directly.
• Mistaking it for a controlnet loader. This node only resizes/enhances the map. The actual ControlNet application happens in a ControlNet Apply-family node downstream. Feed this node's output there.
• Aspect distortion. "Just Resize" stretches. If your pose comes out squashed, switch to "Crop and Resize" or "Resize and Fill" - that's what they're for.
• Video frame counts. It processes every frame in a batch with the same settings, which is what you want for video control maps - just remember it's per-frame, not temporal. Jitter between frames comes from your preprocessor, not this node.

This is a specialist tool. If you run pose/depth/canny-controlled video generation at all, it's the difference between control maps that survive upscaling and ones that dissolve. For plain image gen at fixed resolution you can skip it - but you'll know exactly when you need it.