PixArt Load LoRA

Same job every LoRA loader in ComfyUI does - take a trained set of weight deltas and patch them onto your base model - just wired for PixArt's diffusion-transformer layer names instead of SD or SDXL's UNet ones. Worth setting expectations up front: PixArt never got the deep community LoRA ecosystem SD1.5 and SDXL built up. There was a real moment in 2024 where people floated exactly that - someone specifically imagined "a godlike finetune like PixartPony" pulling the community's attention away from Stability's models - but the ecosystem never fully committed before Flux arrived and reset everyone's priorities. So if you're here, you're more likely training or using a specific PixArt LoRA yourself than browsing a big public library of them.

Inputs and output

• model - a MODEL, from PixArt Checkpoint Loader (auto).
• lora_name - a dropdown of files sitting in your normal ComfyUI/models/loras folder, same location as any other LoRA.
• strength - FLOAT, default 1.0, range -20 to 20, step 0.01. Treat 1.0 as "as trained." If a LoRA is overpowering the base image, nudge it down toward 0.6–0.8 - the same rule of thumb you'd use anywhere else. The huge numeric range is headroom, not a suggestion: push much past roughly 1.5–2 on most LoRAs and coherence starts breaking down regardless of architecture.

Output is MODEL - chain multiple PixArt Load LoRA nodes back to back to stack several LoRAs at once, the same pattern as stock ComfyUI LoRA loaders.

Installing it

Part of the whole ComfyUI_ExtraModels pack:

• ComfyUI Manager - search "Extra Models for ComfyUI", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/city96/ComfyUI_ExtraModels, then pip install -r requirements.txt. Restart.

PixArt LoRA files go in the standard ComfyUI/models/loras folder - no special subfolder needed.

Common issues

This node only patches the MODEL - it doesn't output or touch a CLIP, because PixArt doesn't condition on CLIP at all (it's T5-only). If you're coming from SD1.5/SDXL muscle memory where a LoRA loader hands back both MODEL and CLIP, that's the adjustment: any text-encoder component a LoRA ships won't be applied by this node, so check the LoRA's own documentation if it mentions one.

Because PixArt's ecosystem is thin, it's easy to accidentally load a LoRA that was never trained for this architecture - say, an SD1.5 LoRA someone mislabeled. This node doesn't validate that the weights actually match PixArt's layer shapes; a mismatched LoRA will often load without error and simply do nothing, which can look like the node itself is broken when it's really just a shape mismatch silently no-opping.

If you're training your own PixArt LoRAs rather than downloading one, keep in mind you're on your own more than you would be with SDXL - there isn't a well-trodden community trainer config to copy, so expect some trial and error on rank and learning rate before you get results you're happy with.