This workflow uses three loader nodes connected to the Musubi Wan Trainer node, which handles the full training loop:

• Musubi DiT Loader - loads the Wan 2.2 DiT model (wan2.2_t2v_low_noise_14B)

• Musubi VAE Loader - loads the Wan 2.1 VAE (wan_2.1_vae.safetensors)

• Musubi Text Encoder Loader - loads the T5 text encoder (t5_umt5-xxl-enc-bf16.pth)

All defaults are already optimized for Wan 2.2. Most users only need to do two things before hitting Queue.

Choosing a Task Type

The task setting determines which type of training is performed. The default is t2v-A14B (Wan 2.2 text-to-video), which is the recommended starting point for most users training with image datasets. The trainer supports the following tasks:

• t2v-1.3B - Wan 2.1 text to video (1.3B). Lightweight, lower resource requirements.

• t2v-14B - Wan 2.1 text to video (14B). Higher quality output.

• i2v-14B - Wan 2.1 image to video (14B). Requires CLIP vision encoder and video clip datasets.

• t2i-14B - Wan 2.1 text to image (14B).

• t2v-A14B - Wan 2.2 text to video (14B active, MoE architecture). Default and recommended.

• i2v-A14B - Wan 2.2 image to video (14B active, MoE architecture). Requires video clip datasets.

Important: I2V tasks (i2v-14B and i2v-A14B) require video clip datasets, not standalone images. The I2V training pipeline uses the first frame of each clip as a conditioning image, so a plain image dataset will fail. If you are training with images, use a T2V task instead. T2V LoRAs generally work well for both T2V and I2V inference.

Image Datasets vs. Video Datasets

This guide focuses on image datasets, and we recommend starting there. Image-based training is faster, more cost-effective, and produces excellent results for character and style LoRAs, even though the output model generates video.

Video dataset training requires significantly more compute time and FloTime. A training run that takes 1-2 hours with images can take many hours or even days with video clips, depending on clip count, resolution, and frame length. The costs scale accordingly, and longer training runs leave less room for the iteration that good LoRA training usually requires. Most users find that 2-4 training attempts are needed to dial in the right settings, and that is much more practical at image training speeds and costs.

If you do train with video clips, keep them short (9-25 frames), use a small dataset, and save checkpoints frequently so you can evaluate progress without committing to a full run.

Setting Your Dataset Path

1. Open the Floyo file browser by clicking the folder icon on the left side of the canvas.

2. Navigate to your dataset folder inside #inputs.

3. Click the three-dot menu on your dataset folder and select "Copy Path".

4. Paste the path into the data path field on the trainer node. Your path should look like #inputs/my_dataset. If your dataset lives outside of #inputs, use "Copy path as input" instead, which adds the (as-input) prefix automatically.

For more detailed guidance on the Floyo file browser, see documentation

Name Your LoRA

In the output_name field, type the name you want your LoRA saved as. The trained file will appear in #models/loras when complete.

Key Settings

• task (default: t2v-A14B) - Training task type. See "Choosing a Task Type" above.

• data_path (no default) - Path to your dataset folder (required).

• output_dir (default: #models/loras) - Where your trained LoRA is saved.

• output_name (default: floyo_wan_lora) - Name of your saved LoRA file (change this!).

• resolution (default: 848,480) - Training resolution.

• batch_size (default: 1) - Images processed per training step.

• max_train_epochs (default: 16) - Number of full passes through your dataset.

• save_every_n_epochs (default: 0) - Save a checkpoint each epoch, 0 = only save at the end.

• learning_rate (default: 0.00010) - How fast the model learns, lower is more stable.

• network_dim (default: 16) - LoRA rank, higher captures more detail but creates a larger file.

• discrete_flow_shift (default: 0.0) - Flow matching shift value. Leave at default unless you know what you're doing.

• gradient_checkpointing (default: false) - Trades speed for lower memory usage.

• fp8_base (default: false) - Use fp8 precision for the base model.

• blocks_to_swap (default: 0) - Offload transformer blocks to CPU to save GPU memory.

• target_frames (default: 81) - Number of frames the model targets per training sample.

• frame_extraction (default: head) - How frames are extracted from training data.

Quick Tuning Tips

• Want to compare checkpoints? Set save_every_n_epochs to 1 so you can test each epoch and pick the best one.

• Want more detail captured? Try network_dim at 32. The file will be larger but the LoRA can learn finer features.

• Overfitting or too prompt-rigid? Lower the learning rate to 0.00005 or reduce epochs.

• Training too slow or running out of memory? Enable gradient_checkpointing and try setting blocks_to_swap to offload some work to CPU.