YC Image Accumulator

The usual way to compare generations in ComfyUI is to run a big batch, save everything, and flip through files. YCImageAccumulator is the in-graph version of that habit: it accumulates images across runs, shows you the whole gallery in the node's preview, and lets you pick any one of them to output - or output a slice as a batch. It's a "run 30 seeds, choose the good one" node, and once you've used it you'll miss it in every other workflow.

How it works

The node is stateful - that's the whole point. Every run, whatever image arrives at its image input gets appended to an in-session gallery stored in ComfyUI's temp folder. Only metadata (filename, size, timestamp) stays in memory; full images are written to disk, so it doesn't balloon your VRAM as the gallery grows. selected_index decides what comes out: -1 means the most recent image, 0 the first, anything else that slot. clear_gallery wipes the accumulated images and starts a fresh session.

Because full-res images are saved to a temp folder, the gallery survives across runs within a session - but it is not a permanent store. Flip save_to_output and a copy also lands in your output directory, which is what you want when you finally pick the keeper.

The inputs that matter

• image - the image to accumulate. Anything: VAE output, an upscaler's result, a composite.
• selected_index - output picker, -1 = latest. You'll set this after eyeballing the gallery.
• clear_gallery - set to true once, run, set back to false to start a fresh batch.
• max_images - cap (default 30). When full, new images replace a slot per update_position instead of growing forever.
• batch_start / batch_end - switch from single output to a slice. start=0, end=2 outputs images 0,1,2 as a batch.
• align_mode / pad_color - when outputting a batch of mixed-aspect images, how to align them (first/largest/smallest) and what color pads the mismatches.

Outputs are images and masks (in batch mode the masks mark the content region versus the padding you added with pad_color - that's how you composite mixed-size results back onto one canvas), and it's marked as an output node so the gallery shows in the UI.

How to actually use it

Wire your sampler's output into it, set clear_gallery true for one run to reset, then loop different seeds/denoise/prompts. The node fills up, you scroll the preview, set selected_index to the winner, and that image flows downstream. It shines in workflows where the "best" choice depends on taste you can't automate - character work, composition roulette - and in any A/B where you don't want to re-render to compare.

Installing it

Part of ComfyUI-YCNodes. ComfyUI Manager → search ComfyUI-YCNodes → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/yichengup/ComfyUI-YCNodes

Restart. Only torch/PIL needed.

Gotchas

The statefulness cuts both ways. If you're on a fresh session and the gallery is empty, the node doesn't error - it returns a 64×64 black placeholder with a white mask, which can look like a bug if you don't read the console ([ImageAccumulator] 图像库为空). And because it persists per session, a workflow you load later won't magically restore yesterday's gallery; it starts empty. Also note update_position only matters once max_images is hit - set both, or the "replace slot" behavior never triggers. The node's own console logging ([ImageAccumulator]) is surprisingly chatty and tells you exactly what it's replacing, which makes debugging painless.