Automate diffusion-pipe image or image-edit training with stable LoRA presets and selected full finetune presets. Use when Codex needs to prepare a WSL or Linux diffusion-pipe workspace, enforce fixed dataset paths, generate TOML configs, bootstrap dependencies, or launch VRAM-aware training runs for Flux, Qwen-Image, SDXL, Lumina 2, or HunyuanImage.
Framework-aware workflow for model-engineering tasks in common LLM, VL, and image-generation repositories. Use when Codex needs to inspect or modify training code, config files, launch scripts, model definitions, tokenizer or image-processor settings, checkpoint utilities, PEFT adapters, diffusion pipelines, evaluation harnesses, or experiment notes in PyTorch, Hugging Face Trainer, Accelerate, DeepSpeed, FSDP, PEFT, diffusers, LLaVA-like, Qwen-VL-like, or InternVL-like stacks.