AnyAngle 可选 LoRA · T8

Displayed in the node list as AnyAngle 可选 LoRA · T8 ("optional LoRA"), this is a three-widget LoRA loader with one behavioural quirk that the whole pack's workflow depends on: at strength_model of exactly 0, it returns your model untouched and never looks at the file.

That sounds like a footnote. It's actually the hinge of the design. AnyAngle Studio has two output modes - AnyAngle LoRA mode, where the camera LoRA does the heavy lifting, and base-model mode, where you lean on a structure image (pose, depth, Canny) and Qwen Image 2.1 alone. The Studio node tells the graph which mode it's in by emitting anyangle_lora_strength as 1.0 or 0.0, and the example workflows wire that straight into strength_model here. One graph, both modes, no rewiring and no toggle.

How it works

The Python is nine lines and it isn't hiding anything:

• strength_model == 0 → return (model,). No file lookup, no patch, no error if QI2.1_AnyAngle.safetensors isn't on disk.
• Anything else → it delegates to ComfyUI's own LoraLoaderModelOnly with your lora_name and strength, and raises AnyAngle LoRA filename is empty if the string is blank.

So it is LoraLoaderModelOnly, with a pass-through in front. The practical payoff: if you only ever run base-model mode, you never download the AnyAngle LoRA at all - python install_assets.py without the --download-lora flag is the documented no-LoRA install.

It's model-only on purpose. There's no CLIP input or output, so your Qwen3-VL text encoder passes by untouched, which is correct here - the AnyAngle LoRA is trained against the transformer. Don't try to use it as a general-purpose loader for a LoRA that needs to touch the text encoder; use the regular LoRA Loader for that.

Inputs and output

• model - MODEL in, from your UNETLoader.
• lora_name - STRING, defaults to QI2.1_AnyAngle.safetensors, resolved against ComfyUI/models/loras/.
• strength_model - FLOAT, default 1, range −100 to 100, step 0.01.
• Output: a single MODEL. Wire it into the KSampler's model input.

The one trap in that widget: zero means "skip the LoRA entirely," but negative values don't. -1 loads the LoRA inverted rather than doing something gentle. If you're hunting for a soft touch, the useful numbers are positive and small, like 0.6.

Installing it

It arrives with the pack - there's nothing separate. ComfyUI Manager, search Comfyui-Qwen-Image-2.1-MultiAngle-T8, or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-Qwen-Image-2.1-MultiAngle-T8.git ComfyUI-AnyAngle-Studio-T8
cd ComfyUI-AnyAngle-Studio-T8
python -m pip install -r requirements.txt
python install_assets.py --download-lora   # drop --download-lora if you only use base-model mode

Gotchas

Old workflows pin the strength. If your graph has an earlier LoraLoaderModelOnly with strength_model hardcoded to 1, clicking "base model" in the workbench changes nothing downstream - the node still patches the LoRA in. The workbench will refuse to apply and tell you the strength is still fixed: connect the Studio's anyangle_lora_strength output to the loader, or delete the loader. Newer graphs ship with this loader already wired correctly.

The workbench also watches for stragglers. It scans the graph for any LoraLoaderModelOnly or AnyAngleOptionalLoRAT8 node whose lora_name matches "anyangle" but whose strength_model isn't fed by the Studio node, and flags it - a helpful nudge for the mixed old/new workflow case above.

Missing file, non-zero strength. Pass-through only happens at zero. At 1.0 with no LoRA on disk you get the standard loader failure, which is a fine reminder that you asked for a LoRA you don't have.