FLUX.2 SDNQ Adapter Apply

Klein is the first Flux 2 line with a real LoRA ecosystem - training tooling landed three days after release, and by mid-2026 there were hundreds of adapters on CivitAI, from realism stacks to the SNOFS line to font-atlas generators. This node is how you use those inside the FLUX.2 SDNQ pack. It's the adapter in "Adapter Apply": you feed it the FLUX2_SDNQ_PIPELINE from the Loader, pick a file, set a strength, and get back a pipeline with that adapter baked in - ready to chain into the next adapter or straight into Text to Image / Image Edit.

The chaining is the point. The tooltip says it plainly: chain multiple nodes to stack multiple adapters. Each Adapter Apply registers one "desired adapter" on the session, and the session resolves them in order when a generation node runs. Klein tolerates stacking better than most models - a widely-circulated realism stack ran three LoRAs at strength 1.0 each on Klein 9B without burning out, where Z-Image melted past two. So don't be shy; stack, but read the caveats below.

How it picks files

• lora_name - a combo list of adapter files discovered in models/loras. Here's a neat detail: the pack also scans k4b and k9b subfolders inside models/loras, so you can keep 4B and 9B adapters separated by size. Klein 4B and 9B LoRAs are not interchangeable - applying a 9B LoRA to a 4B pipeline is a real mess and a common early failure. If the combo shows only <manual>, drop your file in models/loras and refresh.
• lora_source_custom - the manual escape hatch: a full local path to an adapter that isn't in the dropdown.
• strength - −5 to 5, step 0.05. The practical range is 0.4–1.2 for most adapters. Set it to 0 to remove that adapter request - useful when you're A/B-ing a stack and want to disable one without deleting the node.
• pipeline - the FLUX2_SDNQ_PIPELINE input, and the output is the same type, so wiring looks like Loader → Adapter Apply → Adapter Apply → Text to Image.

LoRA vs LoKr

Most Klein adapters are standard LoRA, and those go through diffusers' native adapter path - the same one the training tools produce. The node also handles LoKr (a Kronecker-product LoRA variant) through a custom FLUX.2 adapter path, so a LoKr file won't just silently fail to apply. You generally don't need to care which path is taken - pick the file, set the strength, go.

Install and gotchas

Pack-level install: ComfyUI Manager → FLUX.2 SDNQ, or clone https://github.com/GeneralShan/comfyui-flux2-sdnq into custom_nodes, restart, and ensure sdnq/diffusers/peft are installed (ComfyUI 0.17+, Python 3.12+).

Where people actually get burned:

• Wrong size LoRA. 4B adapter on 9B pipeline (or vice versa) → garbage or subtle corruption. Use the k4b/k9b folders and match the model you loaded.
• Strength 1.0 isn't always right. Realism/skin adapters often want 0.6–0.8; some "fix" adapters (like the Klein consistency ones) want less. The 5.0 ceiling exists for experimentation, not because it's a good idea.
• "The LoRA did nothing." Remember Klein's prompt rules - a good adapter on a tag-soup prompt still underperforms. Also, adapter effects on the 4-step distilled model are weaker than on base models; if a LoRA looks like it's not loading, run 8 steps and confirm the file is actually in models/loras (the combo won't list files it can't see).
• Stacking has a ceiling. Klein tolerates more than most, but the realism-stack threads are full of melted hands and phantom limbs at extreme stacks. Batch 4 seeds, inspect, and drop strength rather than adding a fourth adapter.