Wan 2.2 image to video settings that work on 24 GB
Tested Wan 2.2 14B image to video settings for 24 GB in ComfyUI: fp8 models, LightX2V 4 step LoRAs, 480 x 832 at 81 frames in about 2 minutes, plus fixes.
Wan 2.2 14B is the best open image to video model you can run at home right now. It is also big. The full precision weights won’t fit on a consumer card, and the default 20 step settings take a long time per clip.
The good news: with the fp8 files and the LightX2V 4 step LoRAs, a 5 second portrait clip takes about 2 minutes on an RTX 3090 Ti and peaks at about 16.5 GB of VRAM. These are the exact settings used in workflow 03 of the AI Character Kit, and this guide explains each one so you can tune them yourself.
Why does Wan 2.2 use two models?
Wan 2.2 14B image to video is split into two experts:
- A high noise model handles the early steps, where the overall motion and layout are decided.
- A low noise model handles the later steps, where detail and texture are refined.
In ComfyUI you load both and run them one after the other with two KSamplerAdvanced nodes. The first sampler stops partway through and passes the noisy latent to the second. Only one model needs to sit in VRAM at a time, which is why this fits on 24 GB.
Which files do you need?
All from the official Comfy-Org Wan 2.2 repackaged page, except the VAE, which comes from the Wan 2.1 repackaged page.
| File | Folder |
|---|---|
| wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors | diffusion_models |
| wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors | diffusion_models |
| wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors | loras |
| wan2.2_i2v_lightx2v_4steps_lora_v1_low_noise.safetensors | loras |
| umt5_xxl_fp8_e4m3fn_scaled.safetensors | text_encoders |
| wan_2.1_vae.safetensors | vae |
Together that is roughly 33 GB of disk. The original model card is at Wan-AI/Wan2.2-I2V-A14B, and the ComfyUI Wan 2.2 tutorial has the official template if you want a reference.
You also want 32 GB of system RAM. With 16 GB, video tends to be very slow or crash while the models swap in and out.
What are the settings that work on 24 GB?
Here is the full set, then the reasons.
| Setting | Value |
|---|---|
| Size | 480 x 832 (portrait) |
| Length | 81 frames (5 s at 16 fps) |
| Total steps | 4 (2 high noise + 2 low noise) |
| CFG | 1 |
| Sampler / scheduler | euler / simple |
| ModelSamplingSD3 shift | 5 |
| LoRAs | LightX2V 4 step, high on the high model, low on the low model, strength 1 |
| Text encoder | umt5_xxl fp8, CLIP type wan |
How should the two samplers be set?
- High noise KSamplerAdvanced: add noise enabled, steps 4, CFG 1, start at step 0, end at step 2, return with leftover noise enabled.
- Low noise KSamplerAdvanced: add noise disabled, steps 4, CFG 1, start at step 2, end at 10000 (meaning “to the end”), return with leftover noise disabled.
Both use euler with the simple scheduler. The WanImageToVideo node takes your image, the width, the height and the length (81).
What if you turn the LoRAs off?
Without the LightX2V LoRAs, use the Wan defaults instead: 20 steps, CFG 3.5, and split at step 10. Quality is a little richer, but each clip takes many times longer. For social content the 4 step version is the better trade.
Why 480 x 832?
It is the speed setting. It is portrait, so it fits phone content, and it keeps a clip at about 2 minutes. You can raise it to 576 x 1024 if you have time and VRAM, but it is often better to generate at 480p and upscale the finished clip to 1080 x 1920 in CapCut or any video upscaler. The kit’s sample ad was made that way.
Match the size to your image. Portrait image, portrait video. If your source is landscape, swap to 832 x 480. A mismatch is one of the main causes of warping.
Why is the negative prompt in Chinese?
It is the official Wan default negative prompt, and the model was trained with it. Leave it as it is. Translating or “improving” it usually makes results worse.
How do you write a motion prompt?
Describe the movement, not the person. The image already tells the model who she is and what she wears. Repeating it wastes the prompt and can fight the image.
A simple formula:
[what she does], [what moves around her], [camera], natural motionExamples that work well:
She turns her head toward the camera and smiles softly, hair moving in the sea
breeze, waves rolling behind her, handheld phone camera, natural motionShe lifts the cup and takes a sip, then smiles at the camera, steam rising,
natural motionSlow dolly in on her face as she turns to look at the camera, soft light, shallow
depth of field, cinematic“Handheld phone camera” adds the slight shake people expect from UGC. “Slow dolly in” gives a more cinematic feel.
Why does the video warp or melt?
Almost always one of three things:
- The motion is too big. A smile, a head turn, hair in the wind and a slow walk look great. Jumping, spinning or dancing in 5 seconds often falls apart. Ask for smaller movements.
- The size doesn’t match the image. Portrait to portrait, landscape to landscape.
- The source image is busy. Lots of hands, overlapping people or tiny faces are hard to animate. Simple, clear images animate best.
Hands that look wrong usually fix themselves with a new seed. Relaxed or hidden hands are easier than complex gestures.
What if you run out of memory?
- Close other GPU apps, including browsers with hardware acceleration.
- Lower the length to 49 frames (about 3 seconds).
- Restart ComfyUI between big workflows to clear VRAM, especially after running an image model.
- The first run is always slower because the models load from disk. Judge speed from the second run.
How do you make clips longer than 5 seconds?
Take the last frame of your clip, run it through again with a new motion prompt, and join the clips in any editor. Keep the motion prompts consistent so the join is invisible. Three passes give you about 15 seconds.
Where should the input image come from?
Any clear image of a fictional character. If you want the same character across many clips, start from one base image and make each scene with an image editor first. How to keep the same face without LoRA training covers that, and the Qwen Image Edit 2511 guide shows the editing step in detail.
If you want your character to copy a specific performance (talking, holding a product, dancing) rather than a described motion, image to video is the wrong tool. Use a driving video with Wan Animate 2 instead.
No 24 GB card? These exact settings run fine on a rented RTX 3090 or 4090. See running ComfyUI on Vast.ai cheaply.
The AI Character Kit includes this workflow with every setting above already in place, a 4 step on and off switch, and ten tested motion prompts.