WAN 2.2 5b WhiteRabbit InterpLoop — v1.1

Author notes from WAN 2.2 5b WhiteRabbit InterpLoop v1.1.json · selected release

Analyze the content of this video frame and write a concise, single-paragraph description of your predictions around what movement takes place throughout the video sequence that follows. Your description should include the details of the character and scene as a whole but only as they related to the movement that occurs in the scene. In addition, make note of the movements of the particles, blinking of the eyes if any, movement of the hair... this is a moment captured in time, and you are describing these few seconds encompassed by the image. Everything that can move, does move - even minute details of the scene. Do not describe "pauses". Don't minimize the motion with words like "slight" or "subtle". Do not use metaphorical language. Your description must be direct and deceisive. Use simple, common language. Be specific, and describe how each detail in the scene moves, but do not be verbose; each word in your description must have purpose. Use the present tense, as if your predictions are coming true as you type them. You will deliver one paragraph without any additional information and without any special characters that format this response, avoid "The image sequence depicts the character" and describe what happens, without saying "the video ....".

Author notes from WAN 2.2 5b WhiteRabbit InterpLoop v1.1.json · selected release

The “Suggested Prompt for Vision Models” was adapted from a workflow by [NRDX (“AIMaster”)](https://www.patreon.com/AIMasterAA). He's also the creator of Live Wallpaper LoRA! See the [original workflow](https://www.patreon.com/posts/135689184) and connect on [Patreon](https://www.patreon.com/AIMasterAA) or [CivitAI](https://civitai.com/user/NRDX).

Author notes from WAN 2.2 5b WhiteRabbit InterpLoop v1.1.json · selected release

## Diffusion Model *Install into models/diffusion_models/* **FP16:** [wan2.2_ti2v_5B_fp16.safetensors](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/blob/main/split_files/diffusion_models/wan2.2_ti2v_5B_fp16.safetensors) **FP8:** [Wan2_2-TI2V-5B_fp8_e5m2_scaled_KJ.safetensors](https://huggingface.co/Kijai/WanVideo_comfy_fp8_scaled/blob/main/TI2V/Wan2_2-TI2V-5B_fp8_e5m2_scaled_KJ.safetensors) ## Recommended LoRA *Install into models/lora/* [Live Wallpaper Style](https://civitai.com/models/1264662) >**LoRA Tip:** If you're using Live Wallpaper Style, try changing *strength* in the **WanVideo Lora Select** between 1 & 0.75 to see how your results change. Higher LoRA strengths skew results towards the style bias it was trained for! ## Text Encoder *Install into models/text_encoders/* **FP16:** [umt5_xxl_fp16.safetensors](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/blob/main/split_files/text_encoders/umt5_xxl_fp16.safetensors) **FP8:** [umt5_xxl_fp8_e4m3fn_scaled.safetensors](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/blob/main/split_files/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors) ## VAE *Install into models/vae/* [wan2.2_vae.safetensors](https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/blob/main/split_files/vae/wan2.2_vae.safetensors) ### Which versions should I get? Lower precision (FP8) runs faster and uses less VRAM, but you’ll take a hit on detail/temporal stability and are more likely to see artifacts. FP16 is heavier but gives the best quality. ### Install Tip To stay organized, it helps to create subfolders inside of the target models/ folder. It will save you from getting confused later as your collection grows. So, for example, you can put your WAN 2.2 models into subfolders "wan2.2/" eg "comfyui/models/vae/wan2.2/

Author notes from WAN 2.2 5b WhiteRabbit InterpLoop v1.1.json · selected release

Set to 12 if you're using WAN 2.1 for some reason

Author notes from WAN 2.2 5b WhiteRabbit InterpLoop v1.1.json · selected release

Gemma-3 is popular among WAN users for getting good starting points for prompts. If you want to run one locally, check out **[the list of available Gemma models](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d)**. You’ll need to research which model is most appropriate for your machine. You can also run Gemma in the cloud. I trust [Featherless.ai](https://featherless.ai) for cloud compute; they have a strict no-logging policy, they’re the most privacy-focused provider I’m aware of, and they charge one upfront monthly cost, which keeps my costs predictable. It helps me out if you use my **[referral link](https://featherless.ai/register?referrer=4Z0BCRPO)** — or you can **[register normally](https://featherless.ai/register)**. I wrote a program for using cloud vision models over API you might find useful: see [CloudInterrogator](https://github.com/Artificial-Sweetener/CloudInterrogator) on GitHub.

Author notes from WAN 2.2 5b WhiteRabbit InterpLoop v1.1.json · selected release

# WAN Vision Prompt — Quickstart Tutorial You *can* write a prompt by hand, but many people feel intimidated by crafting detailed natural-language prompts for **WAN**—or at least want to see examples first. This guide shows a fast, low-friction way to generate a strong **starting prompt** from an image using a vision model. --- ## What you’ll do 1. **Pick a vision model** (hosted or local). 2. **Upload your input image** to the model. 3. **Paste the “Suggested Prompt for Vision Model”** into the chat with that model. 4. **Copy the model’s description of movement** and use it as your WAN seed prompt. --- ## Option A — Use a hosted vision model (fastest) - **[Gemini](https://gemini.google.com/)** or **[ChatGPT](https://chat.openai.com/)** both work well. - Upload your input image and **paste** the “Suggested Prompt for Vision Model.” - Copy the output description and use it in WAN. > **Note on privacy:** Hosted services process your data on their servers. If privacy is critical, review the provider’s data policy or use the local option below. --- ## Option B — Keep the image private (local vision model) If your image is lewd or you simply want complete local control, run a **local vision model**. - Many people report good results with **Gemma-3**. Browse available models here: **[Gemma-3 models](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d)** - You can run models locally with **[Ollama](https://ollama.com/)**, or choose a privacy-focused cloud you control. > **Tip:** Choose the **smallest** model that fits your hardware and latency needs, then scale up if you need more detail or coherence.

Author notes from WAN 2.2 5b WhiteRabbit InterpLoop v1.1.json · selected release

While it was not the first time I'd had the idea, what convinced me to actually try forcing a loop through seam interpolation was a workflow by [AmazingSeek](https://civitai.com/user/AmazingSeek) which itself was based on a workflow by [ekafalain](https://civitai.com/user/ekafalain). In [Ekafalain's workflow](https://civitai.com/models/1426572/wan-21-seamless-loop-workflow-i2v) and [AmazingSeek's after it](https://civitai.com/models/1720535/wan-21-image-to-video-loop-or-workflow), there is a section called "magic stuff" that interpolates over just the last and first frame. It seems like ekafalain programmed it using existing nodes within the comfy graph itself, I can only imagine how tedious that was. That's what made me decide to code custom nodes to take care of the complicated parts. Now, getting a nice, clean interpolation over the first and last frames is simple!

Author notes from WAN 2.2 5b WhiteRabbit InterpLoop v1.1.json · selected release

# Workflow Version 1.1 ## From the Developer ❤️ I hope this workflow is useful to you! I put a lot of hard work into making it, including designing custom nodes to bring it all together and trying to document as much as possible so it is maximally useful to you. ### Links - Have a look at the **[WhiteRabbit repository](https://github.com/Artificial-Sweetener/comfyui-WhiteRabbit)** for node documentation and atomic workflows if you want a better idea of how to build with the custom nodes here or tweak this workflow. - **My Website & Socials**: See my art, poetry, and other dev updates at [artificialsweetener.ai](https://artificialsweetener.ai). - **Buy Me a Coffee**: You can help fuel more projects like this at my [Ko-fi page](https://ko-fi.com/artificial_sweetener). ### This workflow is dedicated to my beloved **Cubby**. - Find her artwork [all over the internet](https://linktr.ee/cubbyai). - She has many excellent LoRA on [CivitAI](https://civitai.com/user/cubby_ai) for you to explore :3

Author notes from WAN 2.2 5b WhiteRabbit InterpLoop v1.1.json · selected release

In both interpolation nodes here, you can raise clear_cache_after_n_frames if your GPU can handle it. "5" is an extremely pessimistic default intended to hopefully avoid running out of memory on destitute cards.

Author notes from WAN 2.2 5b WhiteRabbit InterpLoop v1.1.json · selected release

# Table of Contents 0. COMMON ERRORS 1. FULL CLIP INTERPOLATION 2. INTERPOLATION MULTIPLIER 3. SAMPLE FRAMERATE TO TARGET 4. LoRA STRENGTH 5. DURATION HINT 6. WORKING SIZE 7. DURATION 8. STEPS 9. CFG 10. USE TILED ENCODER 11. ENABLE UPSCALE 12. CHOOSE UPSCALE MODEL 13. UPSCALE TARGET 14. UPSCALE TILE SIZE 15. UPSCALE BATCH SIZE --- Default settings are tuned towards a "pessimistic VRAM situation" to try and get the workflow to work on as many systems as possible. You need to tune it to your specific system to get the most out of the workflow through a little trial and error! This documentation is to help make that process as painless as possible. --- ## 0. COMMON ERRORS People often report an error like... ```AttributeError: type object 'CompiledKernel' has no attribute 'launch_enter_hook'``` I'm not sure what causes this, though my assumption is it has something to do with the WAN Video Nodes. This should fix it for you: 1. Open "🧩 Manager" 2. Click "Install PIP Packages" 3. Install these two, leave the quotes out: "SageAttention", "Triton-Windows". 3.1 Obviously Triton-Windows is only for Windows users. If you get this error on Linux, I would guess the package for Triton is just "Triton". --- ## 1. FULL CLIP INTERPOLATION `WAN 2.2 5b` generates clips at **24fps**. If you want to increase the framerate, you need to interpolate between each frames. You have two ways to do this and there is an advantage to stacking them. - **Roll Clip & Multiply:** This offsets the location of each frame in your batch so that the seam between the first and last frame is now in the middle of the batch. Then, it multiplies your framrate by the `INTERPOLATION MULTIPLIER`. So if your multiplier is 3, you'll end up with a framerate of 72 assuming the default model framerate. Of course, then we set the clip back so that frame 1 = first in batch again. - **Sample Framerate to Target:** This is a straight conversion from any arbitrary framereate to any other framerate. You could turn off Multiply and just do this if you want, but I think by over-shooting our frame target and then downsampling, we give ourselves an opportunity to smooth out issues at the clip seam. --- ## 2. INTERPOLATION MULTIPLIER Number to multiply framerate by with **Roll Clip & Multiply**. If that is disabled, this does nothing. --- ## 3. SAMPLE FRAMERATE TO TARGET Target framerate to interpolate down or up to. If you have this option disabled in **Full Clip Interpolation**, it does nothing. --- ## 4. LoRA Strength Only controls the first LoRA. You can add another LoRA by chaining them in series. --- ## 5. DURATION HINT Length of your video. Anything longer than 3 seconds seems to have a high cost, though `WAN 2.2 5b` is optimized for up to 5 seconds. - Because we are doing some frame cropping and interpolation to get a loop, the final duration of your clip will not be exact, especially if you use "Fancy". --- ## 6. WORKING SIZE Size of your video on the longest edge. The shape of the video is determined by your input image. --- ## 7. DURATION Length of your video. Anything longer than 3 seconds seems to have exponential cost. `WAN 2.2 5b` is optimized for up to 5 seconds. > **Please note:** Because we are doing some frame cropping and interpolation to get a loop, the final duration of your clip will not be exact. --- ## 8. STEPS This is the number of times the sampler denoises the clip. Each step moves along a fixed schedule that decides how much noise to remove. `WAN 2.2 5b` prefers **30 steps**. Going higher can slightly reduce artifacts, but cost (time/VRAM) climbs quickly and returns diminish past ~30. --- ## 9. CFG CFG, or "Guidance Scale" can be understood as "the rate at which the model follows your prompt". Higher values should adhere more to what you've written, lower values allow the model to "improvise". In the context of WAN, higher CFG seems to translate into more motion, but can also cause strange things to appear. Of course, this is highly dependent on your prompt. **5** is a good place to start. --- ## 10. SCHEDULE SHIFT Oversimplified: Think of this as a "motion vs stability" dial. Higher values push toward more motion and variation; lower values keep things steadier and has less of a risk of introducing artifacts. Technical: Transforms the sampler’s noise schedule by remapping each per-step noise level by a non-linear offset that preserves start and end points. Higher values bias sampling toward earlier/noisier timesteps which will focus the sampler on developing low frequency details while lower values biases toward later/cleaner timesteps, trending instead towards high frequency details. --- ## 11. SEED The number used by the noising algorithm to produce the noise we inject into our latents at the start. This is where "randomness" is introduced in the system, so if you lock the seed and change nothing else, you'll get the same result over and over. If you get a result you really like but with small problems, you might be able to lock the seed and make small changes to some other value to bring it in line with your vision. For example, try locking the seed and tweaking your prompt, or CFG, or Shift. --- ## 12. USE TILED ENCODER This is a technique to avoid out of memory errors when encoding the image into latent space. It's slower to have it turned on, but you may need to if you try to push the limits of the resolution you can do. --- ## 13. ENABLE UPSCALE Turns the post-process upscaler **on/off**. When enabled, frames are upscaled **after** generation and initial loop interpolation, but **before** full clip interpolation (multiply, framerate to target) using the model you pick below. Leave it off to save at your **Working Size** (faster export, lower VRAM). - If you’re troubleshooting quality, disable this first so you can judge the raw clip. - Upscaling increases render time and VRAM use in proportion to your **Upscale Target**. --- ## 14. CHOOSE UPSCALE MODEL Select which **upscale model** to use. The dropdown mirrors whatever you have in `models/upscale_models/`. Need more options? Browse and download from [OpenModelDB](https://openmodeldb.info/), drop the files into that folder, then reload for the options to show up here. --- ## 15. UPSCALE TARGET Sets the **longest edge** (in pixels) the upscaler should aim for. - **1280** ≈ 720p long edge - **1920** ≈ 1080p long edge - **3840** ≈ 4K long edge Higher targets cost more VRAM/time. If you want the working resolution, either disable **Enable Upscale** or set the target near your **Working Size**. --- ## 16. UPSCALE TILE SIZE Controls how large each tile is when the upscaler runs. - **Larger tiles** → faster and fewer tile boundaries, but **more VRAM**. - **Smaller tiles** → safer on VRAM, but slower; can increase the chance of subtle tile transitions. If you hit out-of-memory, lower **Tile Size** before changing anything else here. > **Note:** This is separate from **Use Tiled Encoder** above. That one tiles the **VAE encode/decode**; this one tiles the **upscaler** pass. --- ## 17. UPSCALE BATCH SIZE How many tiles the upscaler processes **at once**. - **Higher batch** → better throughput, **higher peak VRAM**. - **Lower batch** → slower, but more stable on limited GPUs. If you’re tight on memory, reduce **Batch Size** first; if OOM persists, then reduce **Tile Size**.