OmniLoRA is a ComfyUI training node by community author cyberterm: connect the diffusion model, text encoder, and VAE on the canvas, fill in the image folder and trigger word to start training; LoRAs are saved to models/loras, and the trained model can be directly connected to KSampler for output verification. The author states that it automatically identifies architectures from the ComfyUI runtime, supporting Flux, Qwen-Image, Wan, SDXL, HunyuanVideo, LTX-Video, etc., and can also train on FP8, INT8, and GGUF quantized base models; VRAM is calculated before training begins, with layers that don't fit loaded from memory on demand. It includes 5 example workflows and a loss curve panel on the canvas, allowing manual checkpoint saving at any time and automatic saving upon interruption.
Why it matters No need to configure a separate training environment; the author states support for FP8, GGUF, and other quantized base models as well as Intel, AMD, and Apple GPUs, with layers moved to memory when VRAM is insufficient.
GitHub stars 17 · Registry downloads 8
License: Commercial use allowed · ComfyUI: Community node · Scan passed · VRAM: Not specified
Node Audit: N1 passed · N2 with warnings · N3 pending manual review (version 1.0.2)
- MEDIUM pickle / torch.load without weights_only
Reasons
- Apache-2.0; the README provides starting parameters for portraits, styles, quick tests, and low VRAM respectively
- The author states no additional pip dependencies are required, using only ComfyUI's built-in torch, safetensors, and Pillow
Usage
- After installing the node, drag omni_lora_basic_workflow.json from example_workflows into ComfyUI, select models in Load Diffusion Model, CLIPLoader, and VAELoader, fill in the image folder path and trigger_word in OmniLoRA Dataset & Prompt Loader, then click Queue Prompt; for 8–12 GB GPUs, the author recommends selecting adafactor as the optimizer
Risks
- Node audit: N1 passed, N2 with warnings (one instance of torch.load already includes weights_only=True in the code), N3 manual review not yet completed; no installation command is provided this issue
- 1.0.2 was just released, and Registry downloads are very low; supported architectures are tested by the author themselves, and training results vary depending on the base model and data