Characters, styles and video: MOSS Image, Krea 2, FLUX 3, FLUX.2 Klein, Qwen Image 2.1, Z-Image, Anima, Fibo 2.0, SDXL (and its finetunes) and MiniMax H3 video, each trained end to end here; more are listed as untested
Straight into ComfyUI: trained LoRAs show up in its LoRA loaders
Simple or Advanced: a recipe sets everything, or set every option yourself

Quick start

Install ZenTrainer from ComfyUI Manager, restart ComfyUI and click the lotus in its top bar.
New run → Simple: pick a recipe, add 15–40 photos, name it. Missing models are downloaded and the photos captioned for you.
Watch it train. The LoRA lands in output/zentrainer/loras/.

Training you can watch

Each run is one page: the newest samples beside the loss, each sample denoising as it renders, and the numbers in a rail on the right.

Pause, sample, save or change the learning rate mid-run
Resume from any checkpoint
Validation loss on a fixed set, and a likeness score for character LoRAs
Try a prompt on any checkpoint
Progression: every prompt across every step (←/→ steps, ↑/↓ prompts)
A queue: line runs up and they train one after another

A run

Datasets

Any folder in ComfyUI's input/ is a dataset: browse it as a grid or a table, edit captions in place.

Auto-caption with Qwen-VL or Gemma, through ComfyUI
Caption sets: long and short side by side; train on either, or a mix
Bulk edits: prefix, suffix, find and replace
Masks: SAM3 finds the subject, or paint one; it trains only where the mask is white

A dataset: browse it, look closer at one picture and its caption, mask what shouldn't train

Set up a run

Pick a model and its files are filled in (base, text encoder, VAE), each installed or one click from downloading. Start from a recipe, ours or one you saved, and the form flags the usual mistakes: no trigger word, a learning rate the optimizer won't like, too few epochs.

Trigger word, and [trigger] in captions and prompts
Optimizers: AdamW, Prodigy, Lion, CAME, schedule-free, 8-bit and more
LoRA or LoKr, masked training, caption dropout
Class preservation: "a woman" stays a woman, not your character
Preview on a turbo build while training the base

New run

Simple mode

Pick a recipe, add photos, name it, train. See each preview land, and try your own prompt at any step. Switch to Advanced in Settings for every option.

Simple mode: pick a recipe, add photos, watch it learn, try it

Apps (experimental)

Test LoRA renders checkpoints × strengths × prompts on a fixed seed; Generate is plain txt2img with your LoRAs. Turn them on in Settings → Experimental pages. → Writing apps

Install

As a node pack: install ZenTrainer from ComfyUI Manager (or clone it into ComfyUI/custom_nodes/), restart ComfyUI and click the lotus, or open /zentrainer/ on its port. Needs ComfyUI 0.41 or newer. Extra optimizers: pip install -r requirements-optional.txt.

With its own pinned ComfyUI: training stays apart from your main install, so updating it or its custom nodes can't break a run. It reads your models where they are.

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

On Windows, install Git for Windows, clone the repository and double-click start-windows.bat.

The first run installs uv, asks where your models are, picks the PyTorch build for your GPU (CUDA or ROCm) and starts; after that it opens a menu. macOS isn't supported yet.

On a rented GPU: a Docker image with the app behind a password. On Runpod it's one click (a referral link: it supports ZenTrainer at no cost to you). Elsewhere → docker/README.md

Run it (zt)

zt is the command line behind the start scripts. From a clone it's scripts/zt; for zt anywhere, uv tool install --editable ./cli.

zt                      # the menu: setup, updates, start, monitor
zt setup --models-from /path/to/your/ComfyUI   # pinned ComfyUI, reading your models
zt start --open         # http://127.0.0.1:8189/zentrainer/
zt monitor              # live dashboard: queue, runs, loss, memory
zt doctor               # what's installed, where models come from
zt update               # pull ZenTrainer, re-sync ComfyUI and the web app

Headless: zt new > spec.json, zt check spec.json, zt train spec.json --follow. zt --help lists the rest. Guides: docs/cli.md, docs/api.md (the HTTP API) and docs/spec.md (every setting).

--models-from takes a ComfyUI install or a models folder; --extra-model-paths takes a yaml
zt attach /path/to/ComfyUI links the engine into a ComfyUI you already have instead
Your work lives in comfyui/input/ (datasets), comfyui/output/zentrainer/loras/ and comfyui/user/zentrainer/ (runs): reinstall with zt update, not by deleting comfyui/
Compiled blocks (off unless turned on, in Memory → Compile blocks): the first run at a new resolution builds its kernels, a few minutes on Linux with NVIDIA and 10–25 on ROCm or Windows. They're cached in ~/.triton, so later runs start fast

How it works

How it works: your models, ComfyUI, ZenTrainer, your LoRA

The training loop is ZenTrainer's own, with ComfyUI as a library rather than workflows. While a run trains it holds ComfyUI's queue, so nothing else lands on the GPU mid-step. More in docs/ARCHITECTURE.md.

engine/   ComfyUI node pack: training harness (ComfyUI as a library), run API, session node
web/      Vue app (zenkit-ui, ZenKit themes), served by ComfyUI at /zentrainer/
tests/    harness tests that need no ComfyUI
scripts/  zt (setup, start, dev, …); devrun.py runs a spec through the harness without the server
docker/   the container image for cloud GPUs (Runpod, …)
docs/     architecture, apps, guides; banner/ holds the README pictures' sources

Develop

zt dev runs ComfyUI and the web app with hot reload at http://localhost:5173/zentrainer/; engine edits need zt restart. comfy.lock pins the ComfyUI commit and torch build it's tested against: move it with zt update --ref <commit|tag>.

# tests (no GPU)
CUDA_VISIBLE_DEVICES= uv run --no-project --with pytest --with torch --with pillow --with numpy \
  --with aiohttp --with safetensors --with av python -m pytest tests

# a spec straight through the harness (see examples/)
comfyui/.venv/bin/python scripts/devrun.py examples/sdxl-quick.json --comfy comfyui

# README pictures: the app's panels (from an install holding only demo content), then the images
uv run --with playwright --with pillow --no-project python docs/banner/shots.py --url http://127.0.0.1:8188 --run <id> --dataset <folder>
uv run --with playwright --with pillow --no-project python docs/banner/render.py

Credits

Code ZenTrainer carries from other projects, and its licences, is listed in THIRD_PARTY_NOTICES.md.

Faces: InsightFace's models (non-commercial use only), or for commercial use fal's AuraFace with OpenCV's YuNet. The face pipeline follows InsightFace's and OpenCV's.
ai-toolkit by Ostris: the Automagic v3 and v4 optimizers, the weighted timestep table and the training adapters. Also ideas: sigmoid timesteps, the content/style bias, [trigger] in captions and MOSS Image Edit's picture layout.
Fizgig by Peter Neill: ideas, likeness scoring of samples among them; the measurements behind EMA 0.98, MiniMax H3's adapter blocks 20-49 and the sharpest-frame clip stills; and the W8A16 int8 backward kernel by mabseyuk.
VGL by Bria: Fibo 2.0's caption schema and prompts.
OneTrainer by Nerogar: ideas ZenTrainer follows: compiling the repeated transformer block, the int8 backward on int8 tensor cores, and reading its -masklabel mask files.
@0xDELUXA: AMD support's start: the ROCm PyTorch setup and whole-card VRAM on Windows (#1).
@aksugat: the RDNA3 int8 tile set, swept on an RX 7800 XT.