Nunchaku FLUX.2 Klein LoRA Loader

Klein without LoRAs is a model you respect. Klein with LoRAs is a model you build a workflow around. The LoRA ecosystem is where Klein's community actually lives - realism stacks like Smartphone Snapshot Photo Reality, skin fixers, consistency LoRAs, and the whole SNOFS family - and this is the node that applies them to a Nunchaku-quantized Klein model. If you've tried stock LoRA loaders against a Nunchaku model and watched them do nothing, this is why: quantized Klein doesn't expose the same patch targets, and this pack's loaders speak the Nunchaku adapter's language instead.

Why you'd reach for it

The distilled Klein line is fast and good at editing, but its anatomy hit rate is "retry a lot," and the fix that the community converged on is LoRAs plus 8 steps at CFG ~1.2. This node is how those LoRAs get on a quantized model. It sits directly downstream of the pack's NunchakuKleinModelLoader: in goes a MODEL, out comes a patched MODEL you can chain further or send to the KSampler.

How it works

The loader reads the LoRA file from your ComfyUI/models/loras/ folder, validates it against the loaded Klein adapter, and bakes the patch into the model's transformer_options as a spec tuple rather than monkey-patching weights in place. That spec gets re-applied per branch, which is what makes stacking and switching clean. It's profile-aware: it knows the 4B and 9B architectures separately and won't let you silently feed one size's LoRA to the other.

The inputs that matter

Only three, and you only really touch two:

• model - from the pack's Model Loader.
• lora_name - dropdown from your loras folder. This is the one you'll change constantly.
• strength - default 1.0, range -10 to 10. Klein tolerates stacking better than most models, but don't read that as license to run everything at 2.0; 0.7-1.0 per LoRA is the sane starting band. Negative strengths subtract a concept, which is occasionally exactly what you want.

The output is MODEL, wired on to the KSampler.

Installing it

Same pack, same install. Manager → search "Nunchaku-Klein", or:

cd ComfyUI/custom_nodes
git clone https://github.com/tom-m-2020/ComfyUI-Nunchaku-Klein

The real prerequisite is the Vitoom Nunchaku backend (pre-built wheel from huggingface.co/tonera/vitoom-nunchaku) plus tonera's Nunchaku Klein checkpoint in diffusion_models/. No pip dependencies in the pack itself - the wheel is the dependency. LoRAs go in models/loras/ as usual.

Common issues

• LoRAs do nothing - you're almost certainly loading them with a stock ComfyUI LoRA loader instead of this one, or mixing 4B and 9B. They are not interchangeable; a 9B LoRA on a 4B model is noise, and vice versa.
• LoRA won't load / compatibility error - the author's changelog notes LoRA incompatibility fixes in v1.2.1, so if you're on an old version, update the pack (and the backend). Nunchaku LoRA support is still young and version-sensitive.
• Same seed, different image - backend nondeterminism, documented in the README, not your LoRA stack.