Next Trainer

One-click LoRA & full finetune training GUI for Windows — supports Anima / SD 1.5 / SDXL / Flux
Extract and run. No environment setup needed. ~12 GB VRAM for Anima LoRA; Anima full finetune needs ~24 GB.
_{Powered by kohya-ss/sd-scripts, Akegarasu-style GUI.}

中文

Credits

---

_{Home portal — quick links to training, monitor, and onboarding}

---

Get Started in 3 Steps

1. Download  →  SD-Trainer-v2.7.0.7z from [Releases](https://github.com/wochenlong/lora-scripts-next/releases), extract
2. Launch    →  Double-click run_gui.bat (auto-installs deps on first run, ~3 GB)
3. Train     →  Open http://127.0.0.1:28000, pick a model, set params, start training

The portable package ships the default WD tagger wd14-convnextv2-v2 under tagger-models/wd14/ (~400 MB). If Hugging Face download fails, place model.onnx and selected_tags.csv there manually — see docs/tagger-models.md.

Requirements: Windows 10/11, NVIDIA GPU (RTX 20+), ~7 GB disk.

Install from source (Linux / advanced users)

git clone https://github.com/wochenlong/lora-scripts-next.git
cd lora-scripts-next

# Windows
run_gui.bat

# Linux
bash install.bash && bash run_gui.sh

# Optional: install Flash Attention 2 for faster Anima training
# Windows
install_flash_attn.bat
# Linux
bash install_flash_attn.sh

Python 3.10 recommended. See Flash Attention 2 docs for details.

---

What's Supported

Mode	Model / script	Notes
Anima LoRA	LoRA · LoKr · T-LoRA	Flash Attention 2 / xformers / SDPA · from ~12 GB VRAM
Anima LoRA Fast	LoRA only (plugin)	Optional anima_lora runtime · ~16 GB+ · see docs/anima-fast.md
Anima Finetune	Full DiT (anima_train.py)	Sidebar 全量微调 → Anima Finetune · ~24 GB VRAM (4090-class)
SD 1.5 / SDXL LoRA	LoRA · LoHa · LoKr	xformers / SDPA
SD 1.5 / SDXL Finetune	Dreambooth / SDXL finetune	Sidebar 全量微调 → Stable Diffusion
Flux	LoRA	xformers / SDPA

_{Anima LoRA — sidebar, model & dataset form, config preview on the right}

_{Anima LoRA Fast — optional plugin path under 标准模式 / Fast 模式; install runtime from the page before training}

_{Anima Finetune — full DiT weights under 全量微调 in the sidebar}

---

Train Monitor

Automatically opens a monitor page (port 6008) when training starts — GPU stats, training parameters, Loss curves, preview samples, and logs all in one dashboard.

_{GPU load & VRAM, total steps, training params at a glance}

_{Preview samples and TensorBoard-backed Loss / LR curves}

_{Real-time training logs with auto-scroll}

---

VRAM Reference (Anima, 1024 resolution, RTX 4090 benchmarked)

Anima LoRA

VRAM	Configuration	Notes
≥ 24 GB	Default settings	Easiest
≥ 16 GB	gradient_checkpointing	Recommended
≥ 12 GB	Gradient checkpointing	Stable
≥ 10 GB	Gradient checkpointing + blocks_to_swap=16	Slightly slower
≥ 8 GB	Gradient checkpointing + swap 24 + cache TE + LoKr	Tight

Anima full finetune (updates full DiT weights — use Anima Finetune in the WebUI, not LoRA)

VRAM	Configuration	Notes
≥ 24 GB	Default + latents/TE cache	~23–24 GB dedicated VRAM in practice; 4090-class recommended

Documentation

Topic	Link
Anima LoRA Training Guide	docs/anima-training.md
Anima Fast Mode (optional plugin)	docs/anima-fast.md
Open-source notices	NOTICE.md
Anima backend (LoRA + full finetune)	docs/anima-backend.md
Anima full finetune example TOML	docs/examples/anima-full-finetune.toml
Flash Attention 2	docs/flash-attention.md
Train Monitor & SSE API	docs/train-monitor.md
Tagger model directory (tagger-models/)	docs/tagger-models.md
Docker Deployment	docs/docker.md
CLI Arguments	docs/cli-args.md

Changelog

Date	Version
2026-05-28	v2.7.0 — Anima LoRA Fast mode (optional anima_lora plugin): WebUI entry, one-click install, train monitor sync, benchmark docs · see docs/anima-fast.md
2026-05-28	v2.6.0 — Anima full finetune WebUI (anima-finetune), anima_train.py wrapper, 全量微调 nav, train monitor label fix; ~24 GB VRAM reference
2026-05-27	v2.5.3 — Portable dependency health check, sidebar version chip (#54)
2026-05-21	v2.5.0 — UI refresh: new sidebar navigation, home portal page, training monitor dashboard with GPU metrics; CSS cleanup
2026-05-21	v2.4.0 — Training stability: env isolation, NaN filter, sample guard, attn_mode fallback, path normalization; Portable tkinter fix
2026-05-20	v2.3.0 — Train Monitor: TensorBoard-backed curves, parameter checks, log sync
2026-05-19	v2.2.0 — Portable flash-attn fix, crash logging, cross-drive monitor
2026-05-19	v2.1.0 — Flash Attention 2 prebuilt wheels, save-by-steps
2026-05-18	v2.0.0 — First portable release, AMD detection, bf16 fix

Full details in CHANGELOG.md.

Credits

Akegarasu/lora-scripts · kohya-ss/sd-scripts · LyCORIS · T-LoRA — Full attribution in NOTICE.md

---

_{Maintainer: @wochenlong · Contributors}