ComfyUI-T-LoRA

ComfyUI-T-LoRA

Custom ComfyUI loaders for timestep-masked T-LoRA inference.

• Official T-LoRA reference implementation: https://github.com/ControlGenAI/T-LoRA
• Gold-standard training/inference integration reference (multi-model): https://github.com/bghira/SimpleTuner
• Extensive portable LyCORIS implementation (including T-LoRA training): https://github.com/KohakuBlueleaf/LyCORIS

What This Repo Implements

Two inference paths are provided:

• Load T-LoRA (Bypass) / Load T-LoRA (Bypass, Model Only)

  • For official-style attn-processor checkpoints (...to_q_lora.down.weight, etc.)
  • Supports SDXL-style and Flux.1-dev-style official T-LoRA key layouts
  • Injected with ComfyUI bypass adapters
  • Per-step mask applied before UNet forward

• Load T-LoRA (LyCORIS, Bypass) / Load T-LoRA (LyCORIS, Bypass, Model Only)

  • For LyCORIS-style T-LoRA checkpoints (q_layer/p_layer/lambda_layer)
  • LyCORIS modules are parsed, then converted into ComfyUI bypass adapters
  • Per-step mask is set via LyCORIS timestep-mask APIs before each denoise step

Install

ln -s ~/src/ComfyUI-T-LoRA $COMFYUI_PATH/custom_nodes/ComfyUI-T-LoRA

LyCORIS loader nodes require lycoris-lora with T-LoRA support from upstream main:

pip install -U git+https://github.com/KohakuBlueleaf/LyCORIS.git

Restart ComfyUI after updates.

Quick Usage

1. Put base model weights in normal ComfyUI model locations.
2. Put T-LoRA checkpoint in models/loras.
3. Add one loader node:
   • Official checkpoint: Load T-LoRA (Bypass)
   • LyCORIS checkpoint: Load T-LoRA (LyCORIS, Bypass)
4. Feed returned MODEL (and CLIP when using non-model-only node) to your sampler path.

Recommended start params:

• strength_model = 1.0
• max_rank = 0 (auto-infer)
• min_rank = 1
• alpha = 1.0
• max_timestep = 0 (auto-infer)

Debugging Mask Behavior

Set:

• debug = true
• debug_every = 1

You should see step logs with computed active rank.
For adapter-level logs, the first adapter call per step reports active dimensions.

<details> <summary>How Timestep Masking Is Applied</summary>

At each denoise step, the node computes an active rank:

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

Then a binary mask activates ranks [0:r) and suppresses higher ranks for that step.

• Official path: mask is consumed by custom T-LoRA bypass adapters.
• LyCORIS path: mask is set with LyCORIS set_timestep_mask(...) and consumed by converted adapters.

</details> <details> <summary>Checkpoint Format Notes</summary>

Official node expects keys like:

• ...attn1.processor.to_q_lora.down.weight
• ...attn1.processor.to_q_lora.up.weight

LyCORIS node expects T-LoRA module keys (detected by lambda_layer / q_layer / p_layer entries).
Non-T-LoRA LyCORIS modules in a mixed checkpoint are skipped.

</details>

Current Limitations

• strength_clip is currently ignored for these T-LoRA loaders.
• This repo is inference-focused; training is out of scope.
• For broad multi-architecture training workflows, use upstream trainer repos directly (SimpleTuner / LyCORIS).