LoRA Stacker V3

This is LoRA Stacker V2 with one genuinely useful addition bolted on: toggles. If you've used rgthree's Power Lora Loader elsewhere in the ecosystem, the pitch here will feel familiar - that node's most-cited feature is exactly this, per-entry on/off switches so you can flip individual LoRAs on and off without disconnecting anything or hunting down which slot had which file. V3 brings the same idea into a plain MODEL+CLIP stacker.

Same base job as V2: up to ten LoRAs applied to any standard checkpoint (SDXL, FLUX, WAN 2.2, whatever your model/clip inputs came from), using ComfyUI's stock load_lora_for_models under the hood. What's new is control at two levels. toggle_all is the master switch - a single boolean, default on - and when you flip it off, every LoRA in the stack is bypassed at once, outputs pass through unchanged. That's your one-click "show me the base model" comparison without deleting or rewiring anything. Below that, each slot has its own enabled_X boolean (default true), which only matters while toggle_all is on - think of it as the master breaker plus ten individual switches downstream of it. The behavior table is simple: toggle_all off means nothing applies regardless of individual switches; toggle_all on with a slot's enabled_X off skips just that slot; toggle_all and enabled_X both on applies that slot, provided you've actually picked a LoRA file and the strength isn't zero.

The rest of the slot fields match V2: lora_name_X (optional, None to leave empty) and lora_strength_X (optional, default 1.0, range -100 to 100, step 0.01 - one shared strength affecting both model and CLIP together, same as V2, no separate split). There's also a lora_count field, but it's a backend value synced from the frontend's slot-count dropdown rather than something you'd set directly - it caps how many of the ten slots are actually processed. model and clip are required, same as V2, straight from your checkpoint loader.

This is the node to reach for when you're testing whether a LoRA is actually contributing anything, or running a quick style-vs-no-style comparison mid-session - toggle it off, regenerate on the same seed, toggle it back on, compare. It's meaningfully faster than the alternative of deleting a LoRA slot's filename and re-adding it later, and it doesn't cost you the strength value you'd already dialed in.

Outputs: MODEL and CLIP, same as V2 - wire both downstream.

Installing it: bundled with the pack - through ComfyUI Manager, search "ComfyUI-NunchakuFluxLoraStacker," or:

cd ComfyUI/custom_nodes
git clone https://github.com/ussoewwin/ComfyUI-NunchakuFluxLoraStacker.git

Restart ComfyUI, look under loaders.

Troubleshooting: the toggle logic is worth double-checking before you conclude a LoRA "isn't working" - if toggle_all is off, every enabled flag underneath it is irrelevant, and that's an easy thing to forget you set five minutes ago while chasing a different problem. Beyond the toggle mechanics, the usual LoRA gotchas still apply here exactly as they do on V2: architecture mismatch (a LoRA trained for one base doesn't transfer to another), missing trigger words, and the fact that "1.0 is too strong" is SDXL-era advice that doesn't hold on every newer base model - check what the LoRA's own page recommends rather than defaulting to 0.7 out of habit.