Even though I released an article with Wan 2.2 parameters last week, I have been tinkering and experimenting with the Wan 2.2 Lightning lora that have been released and comparing them with the older Wan 2.1 Lightx2v lora. I've also been experimenting with the new Qwen-Image model and have found a set of good parameters for high quality text-to-image using Qwen-Image as a base model and Wan 2.2 Low as a refiner model.
One of the biggest improvement I've found is not related to the lora. The official Wan workflow suggests to use Euler Sampler and Simpler Scheduler, but I have found that Simpler and Normal Scheduler tend to introduce noise to moving elements. The Beta Scheduler seems to perform much better. And, because I am using 1 CFG in these workflows, I find that Res_multistep Sampler works well and runs quickly.
Other changes include using Lightx2v lora instead of Wan 2.2 Lightning lora and changing the FPS to 16.
Resources
These are links to all of the models and loras used in this process.
Models
Kijai Wan 2.2 t2v High FP8
Wan 2.2 t2v High Noise FP8 Scaled
Kijai Wan 2.2 t2v Low FP8
Wan 2.2 t2v Low Noise FP8 Scaled
Kijai Wan 2.2 i2v High FP8
Wan 2.2 i2v High Noise FP8 Scaled
Kijai Wan 2.2 i2v Low FP8
Wan 2.2 i2v Low Noise FP8 Scaled
Qwen-Image FP8
Qwen-Image fp8 e4m3fn
Loras
Lightx2v t2v Rank 256 Lora
Wan 2.1 14b t2v Lightx2v Rank 256 Lora
Lightx2v i2v Rank 256 Lora
Wan 2.1 14b i2v Lightx2v Rank 256 Lora
Lightning t2v LOW v1.1
Wan 2.2 14b t2v Low Noise Lightning 1.1 Lora
SwarmUI Presets
I have included presets for SwarmUI in this article that you can download and apply. That being said, you will have to modify the models and loras with your actual model/lora locations if you use these presets, as they're almost certainly different.
Wan 2.2 Image-to-Video Parameters
Average Gen Time: 3.25 sec per frame, 65 sec per 20 frames (~1 video second)
Display Advanced Options?: Checked
Sampling
Sampler: Res_multistep
Scheduler: Beta
Resolution
480x480
Note: Lower resolutions may be blurry, high resolutions will increase generation time
Image To Video
Video Model: Wan 2.2 I2V A14B High Noise
Video Swap Model: Wan 2.2 I2V A14B Low Noise
Swap Percent: 0.5
Video Steps: 8
Video CFG: 1
Video Resolution: Image
ComfyUI Note: In a ComfyUI workflow, the Video Model is used in the first KSampler and the Video Swap Model is used in the second KSampler. Also, the first KSampler's end_at_step is (Video Steps x Swap Percent) and the second KSampler's start_at_step is also (Video Steps x Swap Percent) and the end_at_step is (Video Steps).
Advanced Video
FPS: 16
Advanced Sampling
Sigma Shift: 8
Video Lora (1st Sampler in ComfyUI)
Lightx2v I2V 14B Rank 256: 2.0 Strength
Video Swap Lora (2nd Sampler in ComfyUI)
Lightx2v I2V 14B Rank 256: 0.5 Strength
Note: I haven't noticed a difference in lora ranks
SwarmUI Note: To save these lora as a preset, include the lora in the preset's Prompt field as: <video><lora:lightx2v_i2v_14b_rank256:2.0> <videoswap><lora:lightx2v_i2v_14b_rank256:0.5> {value}
Wan 2.2 Text-to-Video Parameters
Average Gen Time: 3.25 sec per frame, 65 sec per 20 frames (~1 video second)
Display Advanced Options?: Checked
Model
Model: Wan 2.2 T2V High Noise 14B
Core Parameters
Steps: 8
CFG: 1
Sampling
Sampler: Res_multistep
Scheduler: Beta
Resolution
320x320
Note: Lower resolutions may be blurry, high resolutions will increase generation time
Text to Video
FPS: 16
Note: For longer videos (10 sec+), you can use 8 FPS with 2 Interpolation. But the video may appear somewhat slomo.
Refine/Upscale
Refiner Control Percentage: 0.5
Refiner Method: Step-Swap (SDXL Refiner Model Original)
Refiner Upscale: 1.5
Refiner Model: Wan 2.2 T2V Low Noise 14B
Refiner CFG: 1
ComfyUI Note: In a ComfyUI workflow, the Model is used in the first KSampler and the Refiner Model is used in the second KSampler. Also, the first KSampler's end_at_step is (Steps x Refiner Control Percent) and the second KSampler's start_at_step is also (Steps x Refiner Control Percent) and the end_at_step is (Steps).
Advanced Sampling
Sigma Shift: 8
Base Lora (1st Sampler in ComfyUI)
Lightx2v T2V 14B Rank 256: 2.0 Strength
Refiner Lora (2nd Sampler in ComfyUI)
Lightx2v T2V 14B Rank 256: 0.5 Strength
Note: I haven't noticed a difference in lora ranks
SwarmUI Note: To save these lora as a preset, include the lora in the preset's Prompt field as: <base><lora:lightx2v_t2v_14b_rank256:2.0> <refiner><lora:lightx2v_t2v_14b_rank256:0.5> {value}
Wan 2.2 Text-to-Image Parameters
Average Gen Time: 53 sec
Display Advanced Options?: Checked
Model
Model: Wan 2.2 T2V High Noise 14B
Core Parameters
Steps: 8
CFG: 1
Sampling
Sampler: Res_multistep
Scheduler: Beta
Resolution
360x360
Note: Lower resolutions may be blurry. Greater resolutions may take much longer to generate.
Text to Video
Frames: 1
Refine/Upscale
Refiner Control Percentage: 0.5
Refiner Method: Step-Swap (SDXL Refiner Model Original)
Refiner Upscale: 2
Refiner Model: Wan 2.2 T2V Low Noise 14B
Refiner CFG: 1
ComfyUI Note: In a ComfyUI workflow, the Model is used in the first KSampler and the Refiner Model is used in the second KSampler. Also, the first KSampler's end_at_step is (Steps x Refiner Control Percent) and the second KSampler's start_at_step is also (Steps x Refiner Control Percent) and the end_at_step is (Steps).
Base Lora (1st Sampler in ComfyUI)
Lightx2v T2V 14B Rank 256: 3.0 Strength
Refiner Lora (2nd Sampler in ComfyUI)
Lightx2v T2V 14B Rank 256: 1.0 Strength
Note: I haven't noticed a difference in lora ranks
SwarmUI Note: To save these lora as a preset, include the lora in the preset's Prompt field as: <base><lora:lightx2v_t2v_14b_rank256:3.0> <refiner><lora:lightx2v_t2v_14b_rank256> {value}
Advanced Sampling
Sigma Shift: 1
Qwen-Image w/ Wan 2.2 Text-to-Image Parameters
Average Gen Time: 3.25 sec per frame, 65 sec per 20 frames (~1 video second)
Display Advanced Options?: Checked
Model
Model: Qwen-Image fp8
Core Parameters
Steps: 16
CFG: 3
Sampling
Sampler: Res_multistep
Scheduler: Beta
Resolution
640x640
Note: Lower resolutions may be blurry, high resolutions will increase generation time
Refine/Upscale
Refiner Control Percentage: 0.35
Refiner Method: Step-Swap (SDXL Refiner Model Original)
Refiner Upscale: 2
Refiner Model: Wan 2.2 T2V Low Noise 14B
Refiner CFG: 1
ComfyUI Note: In a ComfyUI workflow, the Model is used in the first KSampler and the Refiner Model is used in the second KSampler. Also, the first KSampler's end_at_step is (Steps x Refiner Control Percent) and the second KSampler's start_at_step is also (Steps x Refiner Control Percent) and the end_at_step is (Steps).
Advanced Sampling
Sigma Shift: 3.1
Refiner Lora (2nd Sampler in ComfyUI)
Lightning T2V LOW 1.1: 1.0 Strength
Note: I haven't noticed a difference in lora ranks
SwarmUI Note: To save these lora as a preset, include the lora in the preset's Prompt field as: <refiner><lora:wan22-lightning_t2v-v11_LOW> {value}
