A better workflow @ https://civarchive.com/models/1719863/wan-21-i2v-nativeorgguf-self-forcinglightx2v-singledual-sampling

This workflow contains:

1. Stack Lora Loader with Self-Forcing & Reward Loras for Performance - Optional -

   (May hinder movements of some Loras, you can toggle off or use different loras to mix and match with different strength. Or run the wf as it is without those loras and adjust steps and CFG accordingly.)

2. Interpolation for Smoothing

3. Simple Video Speed Control

4. Auto input for numbers of frame and frame rate in accordance to speed, length and interpolation inputs. So you don't have to calculate.

Added in V2:

1. Workflow for T2V.

2. Frame Compensation Toggling

3. Option to scale image by width and keep proportion (i2v).

4. Block Swap for memory management.

5. Upscaler with dimension control option.

6. Save Last Frame options.

7. Visible connectors to swap for GGUF loaders.

8. Some minor additions and visual adjustment.

I2V: Best to use Images with width and height in the multiple of 16 or it will be crop.

T2V: Do not go beyond 81 frames or the first few frames might look glitchy.

Workflow and settings example are embedded in the video posted.

Frame Compensation Toggling:

Speed inputs examples:

1.5 = 50% faster - 2 = 2x faster - 0.7 = 30% slower

Toggle ON:

• Compensate for length of the video when Speed is adjusted.

• Increase in values will speed up the video making motion look faster.

  (Warning: increasing too much speed will cause it to generate more frames in order to compensate for the final length of the video.

  A 81 frames 5sec vid with 2 speed will become 162 frames 5sec vid)

  And may cause longer generation time and performance issues.

  (looks better with Interpolation Multiplier of 3-4).

• Decrease in values will cause the video to look slow-mo.

  Best not to go below 0.6 and make sure videos are at least 5 seconds or more.

  Decrease in values will cause the video to decreases in generated frames.

  (looks better with Interpolation Multiplier of 4-5).

Toggle OFF:

• Does not compensate for length of the video when Speed is adjusted.

• Increasing too much speed will cause the video to reduce its length.

  A 5sec vid with 2 speed will become 2.5sec vid

• Decrease in values will cause the video to increase the length of the video.

  A 5 sec vid with 0.7 speed will increase the length of the video to ~7 sec.

Other optional addition includes:

• SageAttention

• TeaCache

• Skip Layer Guidance

• Enhance A Video

• RifleXRoPE

Notes:

• Use SageAttention if you have it installed, increases generation speed.

• TeaCache is not need if you are using 4-10 steps

• Skip Layer Guidance only works together with TeaCache (min 6 steps)

• Enhance A Video is optional

• Use RifleXRoPE if you are generating more than 81 frames

Links to all Models and Loras in workflow.

(Advance)

• There are connectors in the workflow for you to construct or add upscalers or extender etc.

  Will require some knowledge in making and understanding Wan2.1 workflow.