Doom LoRA Stack

Doom LoRA Stack is the pack's replacement for the classic one-LoRA-at-a-time LoraLoader chain. Instead of ten chained nodes with ten pairs of strength widgets, it's one node with up to ten slots, each holding a LoRA dropdown plus strength_model and strength_clip values. Add slots with ＋/－ buttons, stack LoRAs, and everything applies in sequence to a single model (and clip) on the way through.

If you only ever load one LoRA, this is a lateral move - LoraLoader does the same job with fewer widgets. The payoff is for the two real use cases: LoRA-heavy stylization stacks (character + style + concept LoRAs are normal in Flux workflows, and the KB's lora-training doc is full of people stacking three-plus), and batch workflows where you want every LoRA applied consistently every run.

How it works

The stacking math is standard load_lora_for_models chaining, with two conveniences layered on top:

• global_strength - a single multiplier applied to every slot's strengths at once. Turn the whole stack to 60% with one widget instead of re-typing five numbers. It's effectively a master volume for your LoRA rig, and it's the first thing to reach for when a stack reads "too much."
• model_only - a boolean (default false) that drops the CLIP application entirely, forcing strength_clip to 0 for all slots. The tooltip tells you why it exists: video models (LTXV, WAN, MiniMax) don't take CLIP LoRAs the way image models do. Flip this for video checkpoints and stop feeding a clip that gets ignored anyway.

Each active slot (1..lora_count) applies lora_name_i at strength_model_i × global_strength and strength_clip_i × global_strength. Set a slot's name to None and it's skipped, so you can leave gaps. The clip input is optional and ignored when model_only is on.

Outputs: model and clip, both with the stack baked in - wire them straight into a Doom sampler (or any sampler, since this outputs standard ComfyUI types).

One implementation nicety: it caches the last-loaded LoRA file, so repeated runs with the same LoRA don't re-read the safetensors off disk - a real speedup when you're seed-farming a heavy stack.

Installing it

cd ComfyUI/custom_nodes
git clone https://github.com/PeterMikhai/Doom_Flux_NodePack

Restart or use ComfyUI Manager. No extra dependencies. Needs a current ComfyUI (V3 API). The README's DoomAI_nodes.git clone line is stale; the live repo is Doom_Flux_NodePack.

Common issues

• LoRAs applied to a video model look wrong. That's the model_only flag sitting there unflipped. Video models want model-only application; turn it on.
• "Which strength do I set?" strength_model is the one that does most of the visible work; strength_clip steers the textual/CLIP influence. On image models, keep them paired; on a stack that's fighting itself, start by dropping global_strength to 0.7 rather than zeroing slots one by one.
• LoRA dropdown groups look odd on Windows - the pack normalizes path separators so subfolders group correctly in the dropdown; if you're seeing backslash soup, it's a known cosmetic thing, and it usually means an older version. Update.
• Strongly negative strengths are allowed (-100..100) - useful for "unapply" tricks, but easy to mis-set. Check a slot's strength before blaming the model.

It's not glamorous, but for LoRA stacking in one place with a master volume and a video-mode switch, it's the pack's "I'd rather this just work" answer.