Klein 9B LoRA Analyzer (TJ)

Klein 9B LoRA Analyzer (TJ) is the largest of the pack's LoRA analyzer family and the one most people will actually hit, because Flux2 Klein 9B is the model the community actually trains on. Its structure is 8 double blocks + 24 single blocks = 32 blocks, and the node's job is to show you what a given LoRA is doing inside each of those blocks, then let you fix it block by block instead of wrestling one global strength.

How it works

The shared analyzer core parses the LoRA's weights and maps every key onto Klein 9B's double/single block layout, regardless of whether the file uses dot, kohya, diffusers, or LoKr key conventions. You get a per-block view with live strength bars, a use_original toggle to compare against the unpatched model, and one-click presets - keep the strong blocks, balance, or zero out the weak ones. Edits land in a JSON block_config the node applies when it patches your model.

The architecture guardrail matters more on the 9B than anywhere else in the family, because the two Klein variants look alike: drop a 4B LoRA into the 9B analyzer (or the reverse) and the node detects the block-count mismatch - 32 vs 25 - and warns you. Given how easy it is to mix these up in a folder, that check is worth the price of admission.

Inputs and outputs that matter

• model - your Klein 9B model.
• lora_name - the LoRA to analyze and apply.
• global_strength - master strength, -5 to 5.
• use_original - A/B toggle.
• block_config - JSON of per-block overrides.
• clip - optional, for LoRAs that patch the text encoder.

Outputs: model (patched - into the sampler), clip, analysis_text, analysis_json.

Install

ComfyUI Manager → Install Custom Nodes → search TJ_NODE, or:

cd ComfyUI/custom_nodes
git clone https://github.com/designloves2/ComfyUI-TJ_NODE

Restart ComfyUI. No extra dependencies. Category: ✨ TJ_Node/Lora Analyzer.

The context that makes this node interesting

Klein 9B has a reputation as a finicky training target - the training community's own notes say it hates small batch sizes and high learning rates - which means the LoRAs people share for it are more uneven than the average. A character LoRA that nails the face but wrecked the hands is usually not "too strong," it's locally too strong, and a flat strength cut bleeds the good regions along with the bad. That's precisely the failure this node is built to fix, and for Klein 9B in particular the community is deep enough that block-level tuning has real legs.

The honest caveat is the same one as its siblings: it's a debugging and curation tool, not a daily driver. If you download a Klein 9B LoRA and just want to use it, a standard loader is simpler and the analyzer's per-block UI will feel like overkill. But if you're training on the 9B, or you keep hitting a LoRA that's almost good, this is the node that turns "almost good" into "fixed" - with a warning attached for free.