ComfyUI-Custom-Scripts
pythongosssss
ComfyUI-Custom-Scripts enhances COMFYUI with features like autocomplete for embeddings, auto-arranging graphs, and custom node functionalities, improving user experience and workflow efficiency.
2892
2026-01-07
Wan LoRA Trainer
API
Floyo
LoRA
Training
Wan
1
147
Wan LoRA training is the process of fine‑tuning Wan video models (like Wan 2.1 or 2.2 T2V/I2V) with small, targeted LoRA adapters so you can add custom characters, styles, or camera/motion behaviors without retraining the full 14B model.
Overview
Wan models support LoRA (Low‑Rank Adaptation), which lets you train tiny add‑on weights on your own data—images or short clips—while keeping the base Wan checkpoint frozen. Typical setups tune separate LoRAs for character identity, visual style, or motion/camera patterns using 10–50 images or 10–30 short videos plus rich text captions that describe appearance, style, or motion. Training is usually run via scripts like train_wan_lora.py or tools such as AI Toolkit, Diffusion-Pipe, or hosted trainers, with configs that set resolution (often 720p), fps (24), clip length (~4–5 seconds), learning rate, LoRA rank, and number of steps.
What you can train
Common Wan LoRA types include:
Character LoRA: learns a specific person, mascot, or avatar so prompts with a trigger word reproduce that identity across new videos.
Style LoRA: focuses on look and feel (for example, anime, film stock, painterly) across varied subjects while leaving content flexible.
Motion / camera LoRA: teaches temporal behaviors such as orbits, pans, dollies, or looping sprite‑like motions from short, consistent clips.
After training, LoRAs are loaded alongside Wan in SwarmUI/ComfyUI or CLI by specifying the LoRA path and weight, then used like “prompt + base model + LoRA” to generate custom videos.
How training is done
A typical training recipe for Wan 2.x LoRA looks like:
Prepare a dataset folder or JSONL with video paths, frame counts, captions, fps, and seconds.
Run a training script with options such as
--resolution 720 --fps 24 --clip_seconds 4 --max_train_steps 10_000–20_000 --learning_rate 1e-4 --lora_rank 32–64, plus bf16, xFormers, and gradient checkpointing for efficiency.Periodically validate by generating sample clips with the partially trained LoRA, adjusting steps or rank if it under‑fits (weak effect) or over‑fits (artifacts, rigidity).
Many users rely on managed trainers (for example, Wan 2.2 LoRA Trainer APIs) that take a ZIP of images or videos plus a few hyperparameters and return high‑noise and low‑noise LoRA checkpoints optimized for different denoising stages, simplifying configuration and speeding up training 5–10× versus manual scripts.
Read more
N
EXTENSIONS
Nodes & Models
Wan LoRA training is the process of fine‑tuning Wan video models (like Wan 2.1 or 2.2 T2V/I2V) with small, targeted LoRA adapters so you can add custom characters, styles, or camera/motion behaviors without retraining the full 14B model.
Overview
Wan models support LoRA (Low‑Rank Adaptation), which lets you train tiny add‑on weights on your own data—images or short clips—while keeping the base Wan checkpoint frozen. Typical setups tune separate LoRAs for character identity, visual style, or motion/camera patterns using 10–50 images or 10–30 short videos plus rich text captions that describe appearance, style, or motion. Training is usually run via scripts like train_wan_lora.py or tools such as AI Toolkit, Diffusion-Pipe, or hosted trainers, with configs that set resolution (often 720p), fps (24), clip length (~4–5 seconds), learning rate, LoRA rank, and number of steps.
What you can train
Common Wan LoRA types include:
Character LoRA: learns a specific person, mascot, or avatar so prompts with a trigger word reproduce that identity across new videos.
Style LoRA: focuses on look and feel (for example, anime, film stock, painterly) across varied subjects while leaving content flexible.
Motion / camera LoRA: teaches temporal behaviors such as orbits, pans, dollies, or looping sprite‑like motions from short, consistent clips.
After training, LoRAs are loaded alongside Wan in SwarmUI/ComfyUI or CLI by specifying the LoRA path and weight, then used like “prompt + base model + LoRA” to generate custom videos.
How training is done
A typical training recipe for Wan 2.x LoRA looks like:
Prepare a dataset folder or JSONL with video paths, frame counts, captions, fps, and seconds.
Run a training script with options such as
--resolution 720 --fps 24 --clip_seconds 4 --max_train_steps 10_000–20_000 --learning_rate 1e-4 --lora_rank 32–64, plus bf16, xFormers, and gradient checkpointing for efficiency.Periodically validate by generating sample clips with the partially trained LoRA, adjusting steps or rank if it under‑fits (weak effect) or over‑fits (artifacts, rigidity).
Many users rely on managed trainers (for example, Wan 2.2 LoRA Trainer APIs) that take a ZIP of images or videos plus a few hyperparameters and return high‑noise and low‑noise LoRA checkpoints optimized for different denoising stages, simplifying configuration and speeding up training 5–10× versus manual scripts.
Read more
N
EXTENSIONS