whereismyai
KohakuMossTrainer
KohakuMossTrainer is an image finetuning framework that trains MOSS-Image models using LyCORIS algorithms and full finetunes.
KohakuMossTrainer is a finetuning framework designed for MOSS-Image-Base and MOSS-Image-SFT models. It supports every LyCORIS algorithm such as LoRA, LoCon, LoHa, DoRA, and OFT, alongside full finetunes using Muon, AdamW, and EMA. The software runs on Windows and Linux across various GPU architectures using TileLang and Triton kernels. It offers configurable memory profiles to fit hardware from 6 GB cards to multi-GPU nodes. Users can train on native-resolution datasets using formats like folders, JSONL, parquet, and webdataset.
KohakuMossTrainer at a glance
Key features
- Adapters and full finetune support
- Packed variable-length batch implementation
- GPU-specific TileLang and Triton kernels
- Precision options including bf16, fp16, and fp8
- Memory profiles from speed to minimal VRAM
- Native-resolution data planners and caches
- Multi-GPU training with DDP and FSDP2
- Exports to LyCORIS, ComfyUI, and diffusers
Use it for
- Finetuning MOSS-Image models on custom datasets
- Training style and subject LoRAs on low-VRAM GPUs
- Running multi-GPU training jobs on Linux nodes
- Exporting trained adapters for ComfyUI
Alternatives to KohakuMossTrainer
- Soup — Fine-tune and align models like Llama-3.1-8B on a 4 GB laptop GPU.
- Finetuning With Sampling — Official implementation of the Finetuning with Sampling research paper from Harvard.
- SimpleTuner — It supports dozens of model families with advanced memory and enterprise features.
- Ultralytics YOLO — Delivers high-speed and high-accuracy results in real-time.
- Nunchaku — It reduces memory usage and speeds up 4-bit diffusion models significantly.
- SpawnDev.ILGPU.ML — Runs in-browser neural network inference using 6 backends without servers
- LoRA Scripts — Provides a complete web interface and scripts for kohya-ss trainer.
- Qwen21 Nsfw Pose Skill — Provides exact author recipes and 136 precise 3D poses for ComfyUI.
- stable-diffusion.cpp — Runs diffusion models efficiently in pure C and C++ with no external dependencies.
- OneTrainer — Supports a wide range of diffusion models and training methods in one interface.
- Diffusers — A modular toolbox for inference and training of diffusion models.
- Moss Image Prompt Skill — Provides structured prompt rules for MOSS-Image models.
All KohakuMossTrainer alternatives