TS Batch Source

If you've ever wanted to caption a folder of images, generate a hundred distinct prompts, or re-render the same workflow against fifty inputs, this is the node you reach for. TS Batch Source turns a folder, a text file, or a plain count into a job list - and because of how ComfyUI's own batch engine works, everything wired below it runs once per item, independently. A hundred captions become a hundred separate model calls, not one conversation that keeps growing until the model forgets the first picture.

How it works

When a node's output is a list, ComfyUI executes the downstream branch once per element. That's the engine behavior this node leans on: item comes out as a list, so the whole branch below iterates, and each iteration is its own run. It's also fingerprint-aware - it hashes the folder's contents (name, modification time, size), so dropping new images in makes it re-run rather than quietly processing yesterday's set from cache.

The modes that matter

• Images in folder - one job per image, in natural order, so img2 comes before img10 and you don't get a surprise reorder at the ten-image mark.
• Lines in text file - one job per non-empty line. This is the mode that closes the loop with TS Batch Write (see below).
• Count only - N numbered iterations, for when you just want to loop a workflow N times.

Set the mode, give it a path (folder or text file), and you're mostly done. The two fields worth knowing beyond that: start_at skips this many items so a run that stopped halfway resumes without redoing the finished part (and keeps the original numbering), and limit caps the run. 0 on limit means no limit.

The outputs

• item - one path, line, or number per job. Wire this downstream.
• index - position in the job list, starting at 0.
• total - how many items the job has; the same value for every item.
• seed - a distinct seed per item, derived from the seed widget.

That seed output is the quiet hero. A seed widget on a model node holds one number for all hundred calls, so a hundred iterations of the same task come back identical. This output hands every item its own derived seed - reproducible from the base value, different from its neighbours. Note it's not just base + i: neighbouring seeds are a poor source of variety, so it hashes the pair instead.

⚠️ Read one_per_run before a long batch

With the default (list) mode, ComfyUI finishes every copy of one node before it starts the next. Two consequences: results reach disk only after the model has done every item, and a node that stamps per-item metadata - like TS Image Prompt Injector - gets overwritten by the last item before the save nodes run, so every picture ends up with the same prompt.

Flip one_per_run on and set ComfyUI's Batch count to the number of jobs. Each queued run is then a full pass through the graph: the picture is generated, stamped, and saved before the next job starts. Slightly slower, and the only correct choice when you want to watch results arrive or need honest per-image metadata.

It emits paths, not pictures - deliberately

A hundred 4K frames passed along as images would sit in the output cache (roughly 10 GB) before the first caption is written. TS Batch Load Image reads them one at a time instead, so peak memory stays at a single frame. Pair the two and you can caption a folder that would otherwise OOM your box.

The loop worth knowing: TS Batch Write in "one line per item" mode writes a text file that this node reads back in "Lines in text file" mode - captions become generation jobs with no conversion in between.

Install & troubleshooting

Ships in comfyui-timesaver (AlexYez). ComfyUI Manager → search Timesaver → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/AlexYez/comfyui-timesaver
cd comfyui-timesaver
python -m pip install -r requirements.txt

Restart ComfyUI fully. No models, no heavy deps for this node. Two failure messages you might actually see: "Mode 'X' needs a path" when the mode needs a folder/file and you left it empty, and "Nothing to do: the job list is empty after start_at=… and limit=…" - usually a start_at that skips past everything. The typical chain is TS Batch Source → TS Batch Load Image → TS Qwen 3 VL → TS Batch Write.