FLUX LoRA Stack

If you run FLUX.2 Klein with any kind of LoRA stack, this is the node to try before you reach for anything else. The FLUX LoRA Stack from capitan01R/Comfyui-flux2klein-Lora-loader loads up to ten LoRAs in sequence, each with its own strength, on/off toggle, and - the part that matters - its own auto-convert flag for diffusers-format files.

Why does the convert flag matter so much? Most Klein LoRAs ship in diffusers format, with separate to_q, to_k, to_v projections per attention layer. FLUX's native architecture stores those as a single fused QKV matrix, and the single blocks fuse attention plus the MLP gate into one linear1 projection. Load a diffusers LoRA without converting and a chunk of the attention weights never reaches the model. The author's table puts it bluntly: without conversion, "most attention weights never reach the model." This pack fuses them at load time, which is the whole reason it exists. That's not a feature a stock loader gives you.

Why you'd reach for it

Klein is unusually tolerant of LoRA stacking - community testing found Klein 9B handling three LoRAs at strength 1.0 each (a total of 3.0) where Z-Image burned out at two. The classic "realism sandwich" of a snapshot-realism LoRA plus a skin LoRA plus an NSFW/anatomy LoRA is a genuinely common Klein workflow. The FLUX LoRA Stack is the clean way to run that: one node between your model loader and your sampler, no daisy-chaining ten LoraLoader nodes. The main rival is rgthree's Power Lora Loader, which is excellent for general stacking but doesn't do the diffusers→native fusion or the per-layer calibration this pack adds.

Inputs and outputs

It's simple: model in, patched MODEL out. Wire it into your sampler's model input. Everything else is optional slots 1 through 10, and each slot has the same three controls:

• lora_N - dropdown of everything in ComfyUI/models/loras. Pick the file here.
• strength_N - per-slot strength, default 1.0, range -20 to 20. You'll live between about 0.5 and 1.2; the wide range is for mixing/negating experiments.
• enabled_N and convert_N - toggles. convert_N defaults on, which is what you want for diffusers-format Klein LoRAs; flip it off for a file that's already native FLUX format.

Slots are applied in order, and a slot is skipped entirely if it's disabled, set to "None", or has strength 0. You can leave a handful of empty slots in place and just fill the ones you use.

Install

Easiest is ComfyUI Manager - search Comfyui-flux2klein-Lora-loader. Or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/capitan01R/Comfyui-flux2klein-Lora-loader.git

Then restart ComfyUI. There are no model downloads and no pip dependencies beyond numpy, which ComfyUI already ships. Your LoRAs just need to live in models/loras so the dropdown sees them. The pack is MIT-licensed and on the Comfy registry.

Where people get burned

Most of the pain is the format, and it's the kind you can't see from the output - you'll just wonder why a LoRA looks weak. If a LoRA was trained with to_q/to_k/to_v naming and you load it without conversion, the weights silently don't land. Keep convert_N on unless you know the file is native FLUX. Also remember the 4B/9B rule: Klein LoRAs are not interchangeable across sizes, so a 4B file in a 9B workflow is a waste of time no matter how good the loader is. And per the author's README, this is built around the Klein 9B layout (8 double blocks + 24 single blocks) - treat anything else as untested territory.