LoopStart_SEGIMAGE
The loop start that takes a real image in
- loop
- image
- IMAGE
LoopStart_SEGIMAGE is the image-carrying version of the loop-start node. Where plain LoopStart marks the top of a fake loop and takes a first_loop boolean, this one takes an actual image as its seed - the masked segment, crop, or previous output that the loop body should start from - and hands that image out to whatever's inside the loop.
ComfyUI graphs can't have real cycles, so loop pairs like this fake iteration: the nodes between a LoopStart and its matching LoopEnd execute repeatedly instead of once. This node is the front door; the image it emits is what each pass works on.
Inputs and output
loop(LOOP) - the marker wire connecting this node to its LoopEnd_SEGIMAGE partner. This is what defines the loop region.image(IMAGE) - the image that enters the loop body. Wire your starting segment or first frame here.- Output:
IMAGE- the image the loop body consumes, ready to feed your first processing node (a KSampler, a detail pass, whatever the loop repeats).
The full pattern: LoopStart_SEGIMAGE hands the image in, the loop body refines it, LoopEnd_SEGIMAGE's send_to_next_loop feeds the refined result back, and next iteration starts from the better version. It's the loopback technique - feed output back as input - with pixels instead of a flag. Classic uses: progressive img2img refinement of a region, or iterating over segments the way meshmesh-io's other pack, mm-comfyui-megamask, works with masks.
Installing and the fine print
Same pack as the rest of the family, so one install covers all four nodes. ComfyUI Manager: search "mm-comfyui-loopback", install, restart. Or:
cd ComfyUI/custom_nodes
git clone https://github.com/meshmesh-io/mm-comfyui-loopback
Restart after, and you're done - no model downloads, no pip dependencies, just graph logic. The fine print is the pack's shared history: it's around two years old, the GitHub repo is currently unreachable, and it's the one that broke ComfyUI's node-search box back in early 2024 (ctor.title.toLowerCase is not a function - node titles handed back as a list). Fine to keep if an old workflow depends on it; not something you'd reach for when building fresh. If you do loop images through it, keep the iteration count small - each full VAE round-trip costs a little quality.
Inputs (2)
| Name | Type | Default | Description |
|---|---|---|---|
| loop | LOOP | — | |
| image | IMAGE | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| IMAGE | IMAGE | — |