This is a FP16 32gb version of my Wan I2V workflow using the FusionX model, which allows you to generate 480p 5sec vids in 140 seconds. The model weight 32g so you will need to use Ultra pro GPU to avoid the model to load in too long. With ultra pro the first generation will be done in about 450 seconds, but the next generation when fully loaded will be 140 seconds.
Key features :
- Full FP16 wan model with FusionX
-140s generation for 5 seconds of a 480p video
- Lora loader
- 32fps interpolation
- Included Lora High Speed Dynamic for real speed video
How to use :
All models are already loaded
Load image, resize to 480x832 or 720x1280 for the best output, but you can try with different ratio.
You can load different loras with the rgthree lora loader, write your prompt in the positive Text Encode. The Negative prompt shouldn't be needed with a cfg of 1 but in case keep this negative preset.
You can change the generation parameters in this group : Shift control how much your video change from the image, this workflow works with less than 4.
10 steps seems to be optimal, you can go as low as 4. CFG of 1 is required, higher will destroy the generation.
Uni_pc seems to be the best sampler, for the scheduler the best ones seems to be beta, simple, and ddim_uniform. One might be better than the others in certain cases.
You can control the video length, since wan generate in 16fps, 5 seconds = 81 frames, 161 = 10 seconds. The workflow will be out of memory around 200 frames.
The video is generated first at 16fps and then interpolated to 32fps.
High Speed Dynamic Lora :
Allows you to generate videos with realistic speed motion if you get slow motion. Strength control how fast motion will be generated, you might need to lower it in case it glitches the motion. using "real speed" in prompt usually helps.