Stable Video Infinity 2 1080p 60 FPS ComfyUI Workflow for Wan 2.2 I2V by Smittie — v6 — v1

Author notes from Smittie_SVI2_1080p_60FPS.json · selected release

# Smittie's Workflow ## Description With this workflow you can generate long videos in 1080p and 60 FPS without color or motion mismatch between the video segments. At least 16 GB of VRAM recommended, but with more quantized versions of the Wan 2.2 diffusion models, you can go even lower. ## Install Instructions ### 1. [Install KJNodes](https://github.com/kijai/ComfyUI-KJNodes) ### 2. [Install Stable Video Infinity 2](https://github.com/vita-epfl/Stable-Video-Infinity) ### 3. Download all needed models ## Video Instructions If you don't know what to do, you can watch the [Stable Video Infinity Tutorial by AI Search](https://www.youtube.com/watch?v=-3DVJu72VhE). He works with another workflow I build upon. ## Acknowledgement Most credits goes to the Stable Video Infinity, Lightning and Wan team, as well as AI Search and darksidewalker. Thank you guys!

Author notes from Smittie_SVI2_1080p_60FPS.json · selected release

**Diffusion model** - Original: [Wan fp8 models](https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/tree/main/I2V) - Recommended: [DaSiWa WAN 2.2 I2V 14B | Lightspeed](https://civitai.com/models/1981116?modelVersionId=2388548) It is a NSFW model, but it can be used for SFW too and has very good quality and performance. **LoRA** - Only needed for original Wan to speed up generation: [Lightx2v models](https://huggingface.co/lightx2v/Wan2.2-Distill-Loras/tree/main) - Always needed: [SVI 2.0 Pro models](https://huggingface.co/Kijai/WanVideo_comfy/tree/main/LoRAs/Stable-Video-Infinity/v2.0) **Text encoder** - Important as other CLIP's might not work: [CLIP Text Encoder](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/tree/main/split_files/text_encoders) **VAE** - wan_2.1_vae is sufficient: [Wan VAE's](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/tree/main/split_files/vae) **Model Storage Location** ``` 📂 ComfyUI/ ├── 📂 models/ │ ├── 📂 diffusion_models/ │ │ └── │ ├── 📂 loras/ │ │ └── │ ├── 📂 text_encoders/ │ │ └── │ └── 📂 vae/ │ └── ```

Author notes from Smittie_SVI2_1080p_60FPS.json · selected release

### Here you can write a general text prompt that can be used for multiple video generation segments. It is just useful to avoid a lot of copy-paste action, but can be ignored, if you write to each segment an individual prompt anyway.

Author notes from Smittie_SVI2_1080p_60FPS.json · selected release

### 1. If you ran out of RAM, deactivate the image upscaler first, then try again. | Source Resolution | Scale By | Final Resolution | |---------------------|---------------|---------------| | 480p | 1.5x | 720p | | 480p | 2.25x | 1080p | | 720p | 1.5x | 1080p | | 720p | 2x | 1440p | ### 2. The RIFE Frame Interpolation creates frames in-between the generated ones. Most AI models run at 16 FPS, this scales the frames by 3x and therefore to 48 FPS. Then the video will be saved with 60 FPS output to speed up the video by 1.25x, as most AI models default speed looks like a slow-motion. | Target FPS | Frame Rate | Thus Speed Up By | |---------------------|---------------|---------------| | 32 | 60 | ~2x | | 40 | 60 | 1.5x | | 48 | 60 | 1.25x | | 60 | 60 | 1 |

Author notes from Smittie_SVI2_1080p_60FPS.json · selected release

### You can enable the LoRA and/or change CFG, Lenght and Text Prompt per segment individually by removing the line from the Video Generation node. ### You can disable a Video Generation Segment (Extension), if you want to shorten the video length.

Author notes from Smittie_SVI2_1080p_60FPS.json · selected release

### You can test and find the best setting by yourself. The following table is for reference. | Description | BasicScheduler Steps | SplitSigmas Step | Length Per Segment | |---------------------|---------------|---------------|---------------| | For not much movement (Best Performance) | 2 | 1 | 33 | | Good Quality, Performance, Prompt adherence | 4 | 2 | 45 | | Fast movement or new objects | 6 | 3 | 45 | | Model Default | 5 | 2 | 81 | ### As the model default is 81 frames and the result is speed up at the end by 1.25x, 101 frames are actually equivalent to 5s video. So with a length of 33 and 3 video generation segments, the final video will be ~5s long.