[Bug] Wan2.2 LightX2V LoRA Merge Fails to Match `cross_attn.k_img` and `cross_attn.v_img` Keys

[xuansu-genmo]

Description

When merging the Wan2.2-I2V-A14B LightX2V distillation LoRA into the Wan2.2-I2V-A14B low-noise model, some LoRA keys related to image conditioning (k_img, v_img, and img_emb) cannot be matched with the base model keys.

The LightX2V LoRA checkpoint contains additional modules such as:

• cross_attn.k_img
• cross_attn.v_img
• cross_attn.norm_k_img
• img_emb.proj

However, the official Wan-AI/Wan2.2-I2V-A14B checkpoint only contains:

• cross_attn.k
• cross_attn.v
• cross_attn.q
• cross_attn.o

without the corresponding image-specific projection modules.

As a result, only part of the LoRA weights are applied during merging.

Steps to Reproduce

1. Download the official Wan2.2-I2V-A14B low-noise model:
   Wan-AI/Wan2.2-I2V-A14B

2. Download the LightX2V 4-step distillation LoRA:
   lightx2v/Wan2.2-Distill-Loras
   wan2.2_i2v_A14b_low_noise_lora_rank64_lightx2v_4step_1022.safetensors

3. Run the LightX2V LoRA conversion/merging tool:
   python converter.py
   --source <Wan2.2-I2V-A14B-low-noise-model>
   --lora_path
   ...

4. Observe that the LoRA loader reports missing model keys for image-specific attention modules.

Expected Result

All LoRA weights in the LightX2V distillation LoRA should be successfully matched and merged into the Wan2.2-I2V-A14B low-noise model.

The expected behavior is that all LoRA updates, including:

• cross_attn.k
• cross_attn.v
• cross_attn.k_img
• cross_attn.v_img
• img_emb.proj

are applied.

Actual Result

Only a subset of LoRA weights are applied.

The converter reports missing keys such as:

Model key not found: blocks.0.cross_attn.k_img.weight
Model key not found: blocks.0.cross_attn.v_img.weight
...
Model key not found: img_emb.proj.1.weight

and:

Found 290 unused LoRA weights - this may indicate key mismatch
Applied 1054 LoRA weight adjustments out of 1262 possible

The missing weights are mainly related to image-conditioning pathways:

• cross_attn.k_img
• cross_attn.v_img
• cross_attn.norm_k_img
• img_emb.proj

Environment Information

• Operating System: Ubuntu 22.04
• Python Version: 3.12
• PyTorch Version: 2.11.0+cu128
• CUDA Version: 12.8
• GPU: NVIDIA A100
• Commit ID: [Please fill in the LightX2V commit ID]

Log Information

Relevant logs:

Loading LoRA from:
wan2.2_i2v_A14b_low_noise_lora_rank64_lightx2v_4step_1022.safetensors

Model key not found:
blocks.23.cross_attn.k_img.weight

Model key not found:
blocks.23.cross_attn.v_img.weight

Model key not found:
blocks.23.cross_attn.norm_k_img.weight

Found 290 unused LoRA weights - this may indicate key mismatch

Applied 1054 LoRA weight adjustments out of 1262 possible

Additional Information

The LightX2V LoRA checkpoint appears to expect an architecture containing separate image-conditioning attention projections:

cross_attn.k_img
cross_attn.v_img
img_emb.proj

However, the official Wan2.2-I2V-A14B checkpoint only exposes:

cross_attn.k
cross_attn.v

with no image-specific projection modules.

It is unclear whether:

1. the LightX2V LoRA was trained on a converted Wan2.2 architecture,
2. the official model requires a different conversion step before applying the LoRA, or
3. additional key mapping logic is required in the LoRA loader.

A comparison between lightx2v/Wan2.2-Official-Models and Wan-AI/Wan2.2-I2V-A14B weight structures may help identify the intended conversion path.