Introduction

Training a LoRA usually means leaving ComfyUI, wrestling with a separate trainer, and hoping the result looks right when you load it back. ZenTrainer brings the whole job inside. It's an open-source ComfyUI node pack from developer Nynxz that adds a training app behind a lotus icon in ComfyUI's top bar. It trains on ComfyUI's own model code, with the models and GPU you already have.

Version 0.4.0 landed on October 30 (with a 0.4.1 container fix minutes later). It's tagged #OpenSourceHiggsfield, it's on the ComfyUI Manager, the license is MIT, and it now runs on NVIDIA (CUDA) or AMD (ROCm) on Linux and Windows.

Ten models, tested end to end

ZenTrainer 0.4 lists ten families as trained and validated in the app. Anything else ComfyUI knows shows up under Other models, marked untested.

  • MOSS Image
  • Krea 2
  • FLUX 3 Image
  • FLUX.2 Klein 9B (new this release)
  • Qwen Image 2.1
  • Z-Image
  • Fibo 2.0
  • Anima from Circlestone Labs (new, with a recipe based on its makers' settings)
  • SDXL and finetunes such as Pony and Illustrious (new)
  • MiniMax H3 video

The repo's validated recipes include real numbers. The Krea 2 character recipe reached a likeness score of 0.78 at 1,500 steps, the MiniMax H3 character recipe hit 0.64 at 1,750 steps, and FLUX.2 Klein 9B in fp8 reached 0.61 to 0.66 by step 250 on 37 photos.

Training you can watch

Each run gets its own page, with the newest samples next to the loss curve and every sample visibly denoising as it renders.

  • Pause, sample, save or change the learning rate mid-run, and resume from any checkpoint
  • Validation loss on a fixed set, plus a likeness score for character LoRAs
  • Try a prompt on any checkpoint, or flip through a Progression grid of every prompt at every step
  • Queue runs so they train one after another
  • Finished LoRAs land in output/zentrainer/loras/ and show up directly in ComfyUI's LoRA loaders

Datasets and settings

Any folder in ComfyUI's input/ works as a dataset. You can browse it as a grid or a table and edit captions in place.

  • Auto-caption with Qwen-VL or Gemma, run through ComfyUI
  • Keep long and short caption sets side by side and train on either one or a mix
  • Masked training: SAM 3 finds the subject, or you paint the mask yourself
  • LoRA or LoKr, a trigger word with [trigger] in captions, class preservation, and optimizers including AdamW, Prodigy, Lion, CAME and schedule-free

Simple mode keeps it beginner-friendly: pick a recipe, add 15 to 40 photos, and name the run. Missing models download and the photos get captioned for you. Switch to Advanced in Settings when you want every knob.

Install

As a node pack, search ZenTrainer in ComfyUI Manager (ComfyUI 0.41 or newer), restart, and click the lotus. To keep training apart from your main install, run it with its own pinned ComfyUI that reads your existing models:

git clone https://github.com/Nynxz/ZenTrainer && cd ZenTrainer && ./start-linux.sh

On Windows, clone the repo and double-click start-windows.bat. There's also a Docker image and a one-click Runpod template for rented GPUs. macOS isn't supported yet.

Worth knowing

  • 0.4.0 fixes a privacy leak: trained LoRA metadata no longer carries your dataset folder name, LoRA folders or captions. If you shared LoRAs from an older version, keep that in mind
  • The API has no login, just like ComfyUI. Be careful if you start ComfyUI with --listen or expose it through a tunnel
  • The face likeness pipeline uses InsightFace models, which are non-commercial. For commercial work, the README points to fal's AuraFace with OpenCV's YuNet
  • Each base model keeps its own license. MiniMax H3's community license excludes the EU, the UK, South Korea and the US

Conclusion

ZenTrainer makes LoRA training feel like part of the creative loop rather than a side quest: watch your character take shape, poke it with a prompt, then drop it straight into your graphs. With Krea 2, FLUX 3 and even H3 video on the tested list, it's a lovely all-in-one bench for local makers.

Want a quick look at what Flux or H3 can do before you train your own style? Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=zentrainer-0-4-train-loras-inside-comfyui-flux-krea-2-minimax-h3

Original Source

Go teach your models something new!

—Aurelia ♡