Load T-LoRA (Bypass, Model Only)

Load T-LoRA (Bypass, Model Only) is the trimmed version of the official-format T-LoRA loader: it takes a MODEL and a T-LoRA file, applies the timestep-masked rank schedule, and hands back a single patched MODEL. No clip input, no CLIP output. For anyone who already loads their text encoder elsewhere - the standard shape of a Flux graph - this is the cleaner node to reach for, and it's the only one of the two official-format variants you need once CLIP isn't part of the loop.

What it's for

The parent node, Load T-LoRA (Bypass), reads official-format T-LoRA checkpoints - keys like ...attn1.processor.to_q_lora.down.weight and ...up.weight - covering both SDXL-style and Flux.1-dev-style layouts. This variant does exactly that job and stops there. It's ideal when you're swapping the model-side adapter while the conditioning graph stays untouched, or when you're composing a T-LoRA with other model patches and the extra CLIP wire would just add clutter. Since the whole pack ignores strength_clip anyway, you're not losing any text-side adjustment by going Model Only.

How it works

Same mechanism as the full node: at each denoise step the active rank is computed from

r = int(((max_timestep - t) / max_timestep)^alpha * (max_rank - min_rank)) + min_rank

and converted into a binary mask over the adapter's rank dimensions, applied before the UNet forward through ComfyUI bypass adapters. Early steps run at low effective rank; late steps approach max_rank. The inputs are lora_name, strength_model (1.0 to start), max_rank (0 = infer from checkpoint), min_rank (1), alpha (1.0), max_timestep (0 = infer from sampling), and debug / debug_every for per-step rank logging. The one MODEL output feeds straight into your sampler.

Installing it

Self-contained - no LyCORIS, no extra pip packages. Via ComfyUI Manager (search ComfyUI-T-LoRA) or:

cd ComfyUI/custom_nodes
git clone https://github.com/bghira/ComfyUI-T-LoRA

Restart ComfyUI, put the checkpoint in models/loras, and you're done.

Troubleshooting

If you get "No T-LoRA adapters were created from this checkpoint," the file isn't in the official key layout - a LyCORIS-format T-LoRA belongs in the LyCORIS loader instead. An error about not being able to infer rank means you should set max_rank manually. And on Flux, some "unmapped key" warnings in the console are normal: the loader intentionally skips single-transformer-block projection keys that don't map into ComfyUI. When in doubt, enable debug and watch the step logs to confirm the rank mask is actually moving.