This lora creates a longform video (257 frames at once) in the style of a fast-cut edit photoshoot with multiple angles. Have a look at the examples to understand what it does.

I very strongly recommend using the workflow I made specifically for this model, otherwise you're not going to get similar results. The workflow helps you create the starting i2v image, which is crucial for the lora to function correctly. The workflow also does two phases: The first phase creates 257 frames at a low resolution, making it easier to run on low Vram and to quickly find a seed you like. The second phase then splits the video into 4 parts, which gets denoised at high resolution. The first phase won't have good likeness to the subject, but the second phase will fix that, especially if you use a high resolution like 1920x1280. > > Click here to download the workflow < <

The quality of your output will depend a lot on the quality of your image inputs!

The lora is trained at 2x speed, which is why interpolation is needed to bring it back to normal speed. The reason for this was to fit more action into the 257 frames, which then become 513 frames. People with low system RAM might have some difficulty processing so many frames at full res. I'm open to suggestions on how to handle this better in the workflow.

Support me so I can make more models, faster: https://ko-fi.com/the_cook

These distillation loras work best with this model:
Wan_2_2_I2V_A14B_HIGH_lightx2v_4step_lora_v1030_rank_64_bf16 on the high noise

lightx2v_I2V_14B_480p_cfg_step_distill_rank256_bf16 on the low noise

Size guide:

Low Resolution 1st phase size options:

192 x 128
384 x 256
480 x 320
576 x 384

High Resolution 2nd phase size options:

768 x 512
960 x 640
1152 x 768
1344 x 896
1536 x 1024
1728 x 1152
1920 x 1280

Higher resolution= better likeness, so see what you can manage to fit into VRAM