Wan2.2-TI2V-5B I2V Generation Failed - Output is filled with meaningless color blocks (Official Model)

[c-zt-p]

🐛 Bug Description

I'm experiencing generation failures when running official Wan2.2-TI2V-5B model for I2V (Image-to-Video) inference. The output video contains meaningless color blocks and light patterns instead of coherent visual content.

Important Notes:

• ✅ Using OFFICIAL Wan2.2-TI2V-5B model from Wan-AI (not LightX2V's lightweight/distilled version)
• ✅ Issue persists BOTH with and without TAE decoder enabled
• ✅ Full official VAE decoder produces the same problematic output

Symptoms:

• Output video shows random color blocks and light patterns
• No meaningful visual content in generated videos
• TI2V mode with image+text constraints produces barely recognizable object outlines

output_lightx2v_wan22_ti2v_i2v_cat.mp4

---

🔧 Environment Information

Hardware

• GPU: NVIDIA RTX 4090 (24GB VRAM)

Software

• Python Version:Python 3.11.14
• PyTorch Version: 2.8.0+cu128
• CUDA Version: 12.8
• Driver Version: | NVIDIA-SMI 535.309.01 Driver Version: 535.309.01 CUDA Version: 12.2 |

---

📋 Configuration

Model Information

• Model Source: OFFICIAL Wan-AI/Wan2.2-TI2V-5B-Diffusers (NOT LightX2V's lightweight version)
• Model Path: Wan-AI/Wan2.2-TI2V-5B-Diffusers
• Model Type: Full 5B parameter model (official weights)

Command Used

LightX2V/scripts/wan22/run_wan22_ti2v_i2v.sh

Test Configuration: /configs/wan22/wan_ti2v_i2v_4090.json

{
    "infer_steps": 50,
    "target_video_length": 81,
    "text_len": 512,
    "target_height": 256,
    "target_width": 256,
    "num_channels_latents": 48,
    "vae_stride": [4, 16, 16],
    "self_attn_1_type": "flash_attn2",
    "cross_attn_1_type": "flash_attn2",
    "cross_attn_2_type": "flash_attn2",
    "sample_guide_scale": 5.0,
    "sample_shift": 5.0,
    "enable_cfg": true,
    "fps": 24,
    "use_image_encoder": false,
    "cpu_offload": true,
    "offload_granularity": "model",
    "t5_cpu_offload": false,
    "vae_cpu_offload": false,
    "vae_offload_cache": true,
    "use_tae": true,
    "tae_path": "./taehv/taew2_2.pth",
    "need_scaled": true
}