abyz22_lamaPreprocessor

Outpainting has a chicken-and-egg problem: you want to extend the image, but the sampler needs something in the new pixels to anchor on, and noise is a terrible anchor. The trick this node runs is to have LaMa pre-fill the empty border with plausible content, then let your sampler refine that guess. abyz22_lamaPreprocessor does exactly that - it pads your image on random sides, runs LaMa over the padded region, and returns a padded image, a latent, and the LaMa-filled result. It's the "Outpainting (LaMa)" feature from this pack's README, and it's also handy as a data-augmentation step for training.

It's the sibling of abyz22_lamaInpaint, both shipping the bundled LaMa implementation in abyz22/image_control. First run downloads the ControlNetLama.pth weights (~200 MB) from lllyasviel's Annotators on Hugging Face.

Inputs that matter

• pixels - the source image (IMAGE).
• vae - needed to encode the padded result.
• Width, Height (each 0–2048, step 8) - the canvas size to pad up to. Both 0 is the special case: no padding, just encode the image as-is with a full noise mask.
• Width_Ratio_min/max, Height_Ratio_min/max (0–1, default 0.5) - how much of the padding goes on each side. The node picks a random split between min and max, so outpainting isn't always "extend right" - set min=max to force an even split, or use the range to vary it per run.
• seed - randomizes the split; same seed reproduces the same composition.

Outputs: LaMa Preprocessed Image (the padded, alpha-masked image for compositing), LaMa Preprocessed Latent (ready for a sampler), and LaMa Image (the LaMa fill as a plain image).

How it works

For each frame it picks a random up/down and left/right padding split, stamps the original into a random-noise canvas, builds a mask where the padding is "fill me," runs the bundled LaMa model on it, and stitches the result. Then it encodes the padded image and applies a noise mask so your KSampler regenerates the whole canvas rather than fighting an awkward seam. The "zoom out = pad, then sample, then crop back" workflow this enables is the modern, model-aware version of the old crop-and-stitch hack.

Install and gotchas

cd ComfyUI/custom_nodes
git clone https://github.com/abyz22/image_control

or ComfyUI Manager → "image_control", restart. First LaMa run downloads weights from Hugging Face. Requirements add openpyxl, pytorch-lightning, kornia; if imports fail at startup, check omegaconf/einops/yaml, which the pack uses but doesn't declare.

The main trap is canvas size: Width/Height are the target canvas, and if they're smaller than your input the math goes sideways - keep them at or above your image dimensions. And remember LaMa's fill is a starting guess, not a final answer; if you sample the latent at full denoise you get a brand-new border, at low denoise you get LaMa's guess with light polish. Low denoise is usually the better look for outpainting.