Shima Multi Saver

Shima Multi Saver is the all-in-one output node, and it's genuinely a lot of node. You feed it an image and it doesn't just save it - it can simultaneously extract lineart, run Canny edge detection, generate a depth map, produce a normal map, pull highlight/shadow masks, quantize a color palette, and save each one, all from a single pass, each output behind its own enable toggle. If you've ever rebuilt the same pipeline of "save PNG, extract edges for training data, generate depth map" by hand, this is the node that collapses it.

The realistic use cases: batch post-processing for dataset or ControlNet training prep, where you need the original plus its edge/depth/normal variants in one folder structure; or production saves where you want the rendered image plus its utility maps archived with consistent naming and embedded metadata. The naming system alone is worth a look - prefix/suffix, project name, collision ID, timestamp, and a filename-order template - which is the difference between "untitled_003.png" chaos and a clean, sortable archive.

A couple of honest notes before you dive in. This is the heaviest node in the pack. The lineart and depth processors auto-download models from HuggingFace on first use (the pack credits the comfyui-nikosis-preprocessors lineage for the lineart/depth logic), and the palette extraction uses scikit-learn's KMeans. If those aren't installed, the affected outputs degrade gracefully - the node checks for optional deps at import and falls back. So: expect a first-run download, and expect the whole feature set only if your Python env has the extras.

How it works

Each save_* toggle gates its own processing pipeline: lineart (sketch extraction with resolution, invert, intensity, and blur controls), canny (low/high threshold), depth (a DepthAnythingV2 model, or a greyscale fallback if the model can't load), normal (Sobel-gradient-based normal map), highlight and shadow (luminance thresholds), and palette (K-means color quantization with hex codes rendered into a grid). save_mode picks between writing to your ComfyUI output folder and a Preview Only mode that writes to the temp folder (cleaned on restart). Metadata goes into the PNG via user_notes, plus filename components like project_name, base_name, collision_id, and timestamp formatting. export_with_workflow embeds the workflow, so the saved files remain shareable/rebuildable the ComfyUI way.

The inputs that matter

• images - the images to save (required).
• save_original - almost always keep this on; it's your actual render.
• The per-output toggles - save_lineart, save_canny, save_depth, save_normal, save_highlight, save_shadow, save_palette - each with its own parameter widgets (thresholds, model choice, resolution) appearing right below it.
• save_mode - "Save to Disk" vs "Preview Only" for dry runs.
• base_folder, project_name, base_name, prefix, suffix, filename_order, separator - the naming/archiving controls.
• user_notes - metadata text embedded into the file.

Also in the main block: fx_in (an optional IMAGE for an externally-processed FX pass) and save_fx (whether to save it). Then the optional group: shima.commonparams (pulls project name, save path, collision ID, timestamp from Shima's Commons), external_project/external_folder/external_collision_id, subfolder_path, and the standard use_commonparams/allow_external_linking toggles.

Outputs

saved_paths (STRING - the paths it wrote), count (INT - how many files), preview_grid (IMAGE - a contact sheet of everything produced), and image (IMAGE - the original, passed through). The grid preview is the fastest way to sanity-check that your lineart came out right before you trust the folder.

How to install it

Part of the Shima pack - ComfyUI Manager, search "Shima", or:

cd ComfyUI/custom_nodes
git clone https://github.com/KDB-USJP/shima_wf.git Shima

then restart. For the full feature set you may need to add scikit-learn (palette), opencv-python (edge/vision helpers), and huggingface_hub (model downloads) into ComfyUI's Python environment. The pack's own requirements.txt is minimal, so treat those as the real dependencies for this node.

Common issues & troubleshooting

Depth/lineart outputs are missing or blank. Either the model download failed or the optional dependency isn't installed - the node logs a fallback. pip install huggingface_hub into ComfyUI's Python and re-run; check the console for the download status.

Palette saves nothing. No scikit-learn → no KMeans → no palette. Install it and restart.

Files land somewhere unexpected. The save root is your ComfyUI output directory, but base_folder, subfolder_path, and commonparams-driven SAVE_PATH all move it. If files vanish into a subfolder you don't recognize, check external_folder and the commonparams save path first.