diffusion-pipe-auto-train

Diffusion Pipe Auto Train

Overview

Use this skill when the user wants a mostly hands-off diffusion-pipe training flow for image or edit datasets. It bundles a WSL-first bootstrap and launcher script, a fixed workspace layout, and conservative presets optimized for stable training on commodity GPUs instead of open-ended hyperparameter hunting.

Version

Current release: v1.0

This v1.0 release focuses on stable automation for preparing configs, enforcing dataset layout, and launching image or image-edit training runs on top of diffusion-pipe.

Planned optimizations after v1.0:

• automatic dataset download
• dataset cleaning and filtering
• automated tagging or caption generation
• automatic dataset organization before training
• a more complete end-to-end pipeline from raw data to training launch

Scope

Supported automatic paths:

• Image LoRA: flux-dev, qwen-image, sdxl, lumina2, hunyuanimage-2.1
• Image full finetune: sdxl, lumina2
• Edit LoRA: flux-kontext, qwen-image-edit

Out of scope for the automatic path:

• Video training
• Native Windows training
• Experimental full finetunes for models that the repo supports but does not document with a stable low-VRAM recipe

If the user explicitly asks for an unsupported automatic path, explain that diffusion-pipe may support it, but this skill keeps the bundled automation on the stable surface.

Quick Start

1. Confirm the runtime.
   • Use Linux or WSL2 only.
   • Do not attempt native Windows training. diffusion-pipe is built around DeepSpeed pipeline parallelism and the upstream repo documents Windows as impractical.
2. Normalize the request into:
   • model
   • train_kind as image or edit
   • mode as lora or full
   • epochs
   • num_gpus
   • gpu_vram_gb
   • required base model paths
3. Read references/presets.md if you need the support matrix, fixed dataset layout, model path requirements, or VRAM floors.
4. Run every bundled command inside Linux or WSL2, including prepare.
5. Bootstrap or update the training repo:

   python scripts/diffusion_pipe_auto.py bootstrap --repo-root /workspace/diffusion-pipe
   

6. Prepare configs and stable workspace directories:

   python scripts/diffusion_pipe_auto.py prepare \
     --workspace-root /workspace/dp-auto \
     --model flux-dev \
     --train-kind image \
     --mode lora \
     --epochs 16 \
     --num-gpus 1 \
     --gpu-vram-gb 24 \
     --model-arg diffusers_path=/models/FLUX.1-dev
   

7. Launch caching and training:

   python scripts/diffusion_pipe_auto.py launch \
     --repo-root /workspace/diffusion-pipe \
     --manifest /workspace/dp-auto/configs/manifests/flux-dev-image-lora.json \
     --cache-first
   

8. Validate:
   • confirm the generated dataset and train TOML files exist
   • confirm the launch command exits successfully
   • confirm a run directory is created under <workspace-root>/runs