Advanced Image Saver (Aesthetic)

ComfyUI's built-in Save Image is fine for the happy path: PNG, numbered, into the output folder. Advanced Image Saver is for when you want the save step to do actual work - filter out the bad generations before they touch disk, pick JPEG vs WebP per run, write into date-stamped folders, and keep metadata you can read back.

The headline feature is the aesthetic filter. Turn on calculate_aesthetic_score, and every image gets scored by the Aesthetic Predictor V2.5 model; anything below aesthetic_threshold never gets saved. That's a "keep the keeper, skip the 200 misfires" button for unattended batch runs. It's optional and off by default - you have to both install the model (pip install aesthetic-predictor-v2-5) and explicitly allow its remote code, which is the right level of friction for something that loads a model into your process. If you'd rather supply scores from elsewhere (say, a separate scoring node), the optional aesthetic_score input lets you do that without loading the built-in predictor at all.

How it works

• Path: output_path parses [time(%Y-%m-%d)] tokens into real dates, and relative paths stay under ComfyUI's output root. Absolute paths outside output require the explicit allow_absolute_output_path opt-in - a sensible guard, given how easy it is to point a save node somewhere you regret.
• Filenames: filename_prefix, filename_delimiter, filename_number_padding (0001 vs 1), and filename_number_start flips between 0001_prefix and prefix_0001. overwrite_mode = prefix_as_filename turns the prefix into a static name.
• Formats: png, jpg/jpeg (with 1–100 quality and dpi), webp (with a lossless_webp toggle), bmp, tiff.
• Metadata: metadata_mode full / minimal / none, plus embed_workflow for ComfyUI-restorable graph data. The honest caveat the README gives you: JPEG metadata support is limited - if you need to read metadata back later, use PNG or WebP.
• Previews: show_previews off for headless runs. If you save to an allowed absolute path outside ComfyUI output, the file still appears in files, just not as a clickable preview (ComfyUI's /view only serves host-managed paths).

Inputs and outputs

The ones you'll set: images, output_path, extension, quality, filename_prefix, and (if you use the filter) calculate_aesthetic_score + aesthetic_threshold. The default threshold is 5 on a 0–10 scale - start there and tune.

Three outputs: filtered_images (images that passed, ready for downstream nodes), files (the saved paths - wire into a text/log node), and scores (per-image aesthetic scores or diagnostics).

Install

The pack plus one optional extra:

cd ComfyUI/custom_nodes
git clone https://github.com/rookiestar28/ComfyUI_Text_Processor.git
pip install -r requirements.txt
# only if you want built-in aesthetic scoring:
pip install aesthetic-predictor-v2-5

Restart after. Manager users: search "ComfyUI Text Processor".

Gotchas

The aesthetic model needs allow_aesthetic_remote_code enabled before it'll load - that's a deliberate trust gate, and it's disabled by default. aesthetic_precision defaults to auto, which picks the best supported device/precision and falls back; if it's misbehaving on a weird setup, cpu_fp32 is the safe choice. And remember: calculate_aesthetic_score off means no filtering - the threshold does nothing. It's a lot of knobs, but the core loop - path, format, filter - is set in three clicks.