nhuquangls/ha379-wan-lora

ha379 Wan 2.2 I2V 14B LoRA

This repo contains the ha379 dataset, AI Toolkit training config, LoRA checkpoints, and optimizer state for continuing training Wan 2.2 I2V 14B.

Trained LoRA

Each LoRA checkpoint is split into Wan 2.2 high-noise and low-noise files. Available checkpoints:

• loras/ha379_wan22_i2v_14b_lora_step5000_high_noise.safetensors
• loras/ha379_wan22_i2v_14b_lora_step5000_low_noise.safetensors
• loras/ha379_wan22_i2v_14b_lora_000008000_high_noise.safetensors
• loras/ha379_wan22_i2v_14b_lora_000008000_low_noise.safetensors
• loras/ha379_wan22_i2v_14b_lora_000008500_high_noise.safetensors
• loras/ha379_wan22_i2v_14b_lora_000008500_low_noise.safetensors
• loras/ha379_wan22_i2v_14b_lora_000009000_high_noise.safetensors
• loras/ha379_wan22_i2v_14b_lora_000009000_low_noise.safetensors
• loras/ha379_wan22_i2v_14b_lora_000009500_high_noise.safetensors
• loras/ha379_wan22_i2v_14b_lora_000009500_low_noise.safetensors

Use both files together in ComfyUI Wan 2.2 I2V workflows. The trigger word is:

ha379

Training Config

Config file:

train_lora_wan22_i2v_14b_ha379.yaml

Important settings:

• Model arch: wan22_14b_i2v
• Base model: ai-toolkit/Wan2.2-I2V-A14B-Diffusers-bf16
• LoRA rank/alpha: 32/32
• steps: 10000
• save_every: 500
• quantize: false
• low_vram: false
• cache_latents_to_disk: false
• disable_sampling: true
• Dataset: dataset/ha379

Notes:

• cache_latents_to_disk is disabled because this Wan I2V training path needs the image tensor at train time. Latent caching caused a NoneType tensor error.
• Sampling is disabled because sample generation failed for this setup around sample time, while the training loop and checkpoint saving work normally.
• Text embeddings are cached to disk and the cache folders are ignored by git.
• The final step 5000 output from AI Toolkit was renamed with an explicit step5000 suffix before upload.
• Training was resumed from step 5000 and stopped after the step 9500 checkpoint. Because max_step_saves_to_keep is 4, only checkpoints 8000, 8500, 9000, and 9500 from the resumed run were kept.

Resume State

Optimizer states:

resume_state/optimizer_step5000.pt
resume_state/optimizer_step9500.pt

To continue training, copy or link the desired LoRA files and matching optimizer state into the AI Toolkit output folder for the job, then run AI Toolkit with the config above. The config is set to train to step 10000. For the latest state, resume from the step 9500 LoRA pair and optimizer_step9500.pt.

Example local paths from the original run:

/root/ai-toolkit/output/ha379_wan22_i2v_14b_lora/
/root/ha379-wan-lora/train_lora_wan22_i2v_14b_ha379.yaml

ComfyUI Placement

Copy the LoRA files into:

ComfyUI/models/loras/

Wan 2.2 I2V also needs the Wan diffusion models, text encoder, and VAE installed in the appropriate ComfyUI model folders.