KSampler Immediate Save

Here's the problem this node solves, and it's a real one: in a big output-list run, ComfyUI waits until every image is processed before anything gets saved and the grid is assembled. A 27-image grid means 27 samplers finishing before the first PNG appears. If you cancel, or your GPU driver blinks, you lose the whole batch. The author calls this "the execution stalling problem," filed an RFC about it upstream, and built this beta node as a stopgap: it saves each image the moment it's done, one file at a time.

It's a node expansion - at runtime it copies the standard CheckpointLoaderSimple → CLIPTextEncode → KSampler → VAEDecode → SaveImage graph, wires your inputs in, and runs it as one unit per invocation. So you get immediate saving without any custom sampler logic. The author is openly ambivalent about it ("A custom KSampler just to save an image? Now I have become the very thing I sought to destroy!"), which should tell you exactly how much of a compromise it is - but it works, and for grid workflows it's a genuine safety net.

Finding it: it's hidden by default

This node lives in the _for_testing category, so it won't appear in search until you flip Settings → Comfy → Show experimental nodes in search on. Do that, and you'll find "KSampler Immediate Save."

The inputs that matter

They mirror the standard KSampler plus a checkpoint and a filename:

• ckpt_name - checkpoint dropdown.
• positive / negative - conditioning text.
• latent_image, seed, steps, cfg, sampler_name, scheduler, denoise - all the usual suspects.
• filename_prefix - save prefix, default ComfyUI; supports %date:...% formatting.

The single image output is the decoded image (not a list) - wire it onward if you still want the grid on top of the immediate saves.

The honest limitations

It only works with the stock pattern. No CFGGuider, no ModelShift, no dual samplers, no custom sampling. If your workflow deviates, the README says plainly: you have to extend the pack's ksampler_immediate_saveimage.py or use a different workaround (the PrimitiveInt control_after_generate=increment pattern, or for-loops). The author would rather you didn't need this node at all - it's explicitly a beta - but if your grid runs take minutes each and you value not re-running 20 of them because you cancelled at image 14, this is the one to grab.

Installing

Search "OutputLists Combiner" in ComfyUI Manager, or clone into ComfyUI/custom_nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/geroldmeisinger/ComfyUI-outputlists-combiner
cd ComfyUI-outputlists-combiner
uv pip install -r requirements.txt

Restart after. The pack recommends running with --cache-ram alongside this so ComfyUI can pick up cached intermediate results after an interruption.