Nodes/Bobs_LoRA_Loader/Bobs LoRA Loader (Universal)
ComfyUI Node

Bobs LoRA Loader (Universal)

Block-weighted LoRA loading for any architecture ComfyUI supports — SD1.5, SD2, SDXL, SD3/3.5, FLUX, Chroma, AuraFlow, PixArt, HiDream, Qwen-Image, Wan, LTX-Video, Mochi, HunyuanVideo/DiT, Lumina, Cosmos and others. The block stacks are discovered from the loaded model itself rather than a hard-coded table, so new and pruned architectures work without an update. Weights are assigned by depth: Early through Late across the whole block stack, plus embeddings, output head and text encoder.

By BobsBlazed·Created about a year ago·Updated 7 days ago· 13
Bobs LoRA Loader (Universal)
  • model
  • clip
  • MODEL
  • CLIP
  • info
lora_name
strength1.00
preset
Text Encoder1.00
Input & Embeddings1.00
Early Blocks (Composition)1.00
Early-Mid Blocks (Subject)1.00
Mid Blocks (Concept & Style)1.00
Late-Mid Blocks (Detail)1.00
Late Blocks (Texture)1.00
Output Head1.00
Other Tensors1.00
CategoryBobs_Nodes

Inputs (14)

NameTypeDefaultDescription
modelMODELDiffusion model to patch.
lora_nameCOMBOLoRA file to load. 'None' passes the model through untouched.
strengthFLOAT1.00-5–5Global multiplier applied on top of every block weight.
presetCOMBOChoose 'Custom' to use the sliders below; any other preset overrides them.
Text EncoderFLOAT1.00-2–2All text-encoder weights (CLIP / T5 / LLM), whatever the model uses.
Input & EmbeddingsFLOAT1.00-2–2Patch, timestep, guidance and context embedding layers feeding the block stack.
Early Blocks (Composition)FLOAT1.00-2–2First 20% of the block stack. Global composition and layout.
Early-Mid Blocks (Subject)FLOAT1.00-2–220–40% of the block stack. Subject identity.
Mid Blocks (Concept & Style)FLOAT1.00-2–240–60% of the block stack. Concept and dominant style.
Late-Mid Blocks (Detail)FLOAT1.00-2–260–80% of the block stack. Detail generation.
Late Blocks (Texture)FLOAT1.00-2–2Final 20% of the block stack. Fine texture and surface.
Output HeadFLOAT1.00-2–2Final projection back to latent space.
Other TensorsFLOAT1.00-2–2Tensors that matched no stack, embedding or head pattern.
clipoptCLIPOptional. Leave unconnected to patch the model only.

Outputs (3)

NameTypeDescription
MODELMODELModel with the block-weighted LoRA applied.
CLIPCLIPCLIP with the text-encoder portion of the LoRA applied.
infoSTRINGPer-block report: weight, tensors found and tensors actually patched.