Wan 2.2 in ComfyUI on 8 to 16 GB (2026): Fast 4-Step Setup, GGUF Files and Best LoRAs

Wan 2.2 is more than a year old, yet in November 2026 it is still the most downloaded open video model on Hugging Face. Comfy-Org’s Wan 2.2 repo alone has passed 9 million downloads. There are three reasons: it is Apache 2.0, it has the biggest LoRA ecosystem of any video model, and its umT5 text encoder never refuses a prompt, which makes it the most creatively open video model you can run. The catch is speed and memory. Out of the box, the 14B model takes 20+ steps on two separate “experts” and wants far more than 16 GB of VRAM. This guide shows the setup that actually works on 8, 12 and 16 GB cards in 2026: the right GGUF files, the LightX2V 4-step LoRAs (including the newer 1022 and 1217 versions and the slow-motion fix), the exact ComfyUI folders, the best settings and the most useful creative LoRAs.

Which Wan 2.2 Model Do You Need?

The A14B models are mixture-of-experts: the high-noise expert lays out the composition and motion in the first steps, then the low-noise expert adds detail. Only one is active at a time, so VRAM is set by one expert, but you need both files. For most people the I2V A14B is the best starting point, because starting from an image (for example one made with Qwen-Image 2.1 or FLUX 3 Klein) gives far more control than text alone.

VRAM Picks

Plan for at least 32 GB of system RAM (64 GB is better): ComfyUI moves the inactive expert and the text encoder into RAM while the other works.

All Files and ComfyUI Folders

The official ComfyUI files are in Comfy-Org/Wan-2.2 (under split_files/). GGUF files come from QuantStack.

Image to video (14B)

• High-noise expert: ComfyUI/models/diffusion_models/wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors, or GGUF ComfyUI/models/unet/Wan2.2-I2V-A14B-HighNoise-Q4_K_M.gguf from QuantStack/Wan2.2-I2V-A14B-GGUF
• Low-noise expert: ComfyUI/models/diffusion_models/wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors, or GGUF ComfyUI/models/unet/Wan2.2-I2V-A14B-LowNoise-Q4_K_M.gguf

Text to video (14B)

• ComfyUI/models/diffusion_models/wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors and ComfyUI/models/diffusion_models/wan2.2_t2v_low_noise_14B_fp8_scaled.safetensors, or GGUF from QuantStack/Wan2.2-T2V-A14B-GGUF in ComfyUI/models/unet/

Shared files

• Text encoder: ComfyUI/models/text_encoders/umt5_xxl_fp8_e4m3fn_scaled.safetensors (or umt5_xxl_fp16.safetensors if you have RAM to spare)
• VAE for the 14B models: ComfyUI/models/vae/wan_2.1_vae.safetensors
• VAE for the 5B model: ComfyUI/models/vae/wan2.2_vae.safetensors

4-step LoRAs

• Official Comfy-Org copies: ComfyUI/models/loras/wan2.2_i2v_lightx2v_4steps_lora_v1_high_noise.safetensors and ..._low_noise.safetensors; T2V: wan2.2_t2v_lightx2v_4steps_lora_v1.1_high_noise.safetensors and ..._low_noise.safetensors
• Newer LightX2V versions from lightx2v/Wan2.2-Distill-Loras: wan2.2_i2v_A14b_high_noise_lora_rank64_lightx2v_4step_1022.safetensors, wan2.2_i2v_A14b_low_noise_lora_rank64_lightx2v_4step_1022.safetensors, wan2.2_t2v_A14b_high_noise_lora_rank64_lightx2v_4step_1217.safetensors and wan2.2_t2v_A14b_low_noise_lora_rank64_lightx2v_4step_1217.safetensors, all in ComfyUI/models/loras/

Not sure where a file goes? Our ComfyUI model folder cheat sheet covers every type. GGUF files can also live in ComfyUI/models/diffusion_models/; the GGUF loader reads both.

Install the Nodes

The native Wan 2.2 workflows need no custom nodes, just an up-to-date ComfyUI (update first, the templates rely on recent core nodes). For GGUF files, install ComfyUI-GGUF by city96:

• ComfyUI Manager: search ComfyUI-GGUF and click Install.
• Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/city96/ComfyUI-GGUF
cd ComfyUI-GGUF
pip install -r requirements.txt

On the Windows portable build run ..\..\..\python_embeded\python.exe -m pip install -r requirements.txt from that folder instead. It installs to ComfyUI/custom_nodes/ComfyUI-GGUF/ and adds the Unet Loader (GGUF) node. For saving MP4s with more options, ComfyUI-VideoHelperSuite (Manager name: ComfyUI-VideoHelperSuite) is handy but optional.

The Fast 4-Step Setup (Step by Step)

1. Open the template. Workflow > Browse Templates > Video > video_wan2_2_14B_i2v (or video_wan2_2_14B_t2v).
2. Swap the loaders for GGUF (16 GB and below). Replace each Load Diffusion Model node with Unet Loader (GGUF) and pick the HighNoise and LowNoise Q4_K_M files.
3. Add the LoRAs. Insert a LoraLoaderModelOnly after each loader: the high-noise 4-step LoRA after the high-noise model, the low-noise LoRA after the low-noise model. Strength 1.0 for both.
4. Set the two samplers. The template uses two KSamplerAdvanced nodes. Set total steps to 4 on both. The high-noise sampler runs steps 0 to 2, the low-noise sampler runs steps 2 to 4. Set CFG to 1.0 on both (the distill LoRAs are trained for CFG 1, and CFG 1 also halves the time per step).
5. Load your image (I2V) and set size and length in the WanImageToVideo node: 832×480 and 81 frames is a safe start (about 5 seconds at 16 fps).
6. Queue. On a 16 GB card, expect the first run to take longer while models load. Later runs are much faster.

The model files stay the same if you later want full quality: remove the LoRAs, set 20 steps (10 + 10) and CFG 3.5.

Best Settings

Want smoother 60 fps output? Run the clip through RIFE interpolation afterwards, as in our RIFE + SeedVR2 workflow; the same nodes work for Wan clips. For sampler theory, see ComfyUI samplers and schedulers explained.

Wan 2.2 5B for 8 GB Cards

The TI2V 5B model does text-to-video and image-to-video in a single 5B file and renders 720p at 24 fps. Comfy’s documentation says it fits 8 GB with ComfyUI’s native offloading. It uses a different, high-compression VAE.

• Model: ComfyUI/models/diffusion_models/wan2.2_ti2v_5B_fp16.safetensors (GGUF versions: QuantStack/Wan2.2-TI2V-5B-GGUF)
• VAE: ComfyUI/models/vae/wan2.2_vae.safetensors (not the 2.1 VAE)
• Text encoder: the same umt5_xxl_fp8_e4m3fn_scaled.safetensors
• Template: video_wan2_2_5B_ti2v. Set length to 121 frames for 5 seconds at 24 fps.

Quality is below the 14B, especially for complex motion, but it is fast and a good way to test prompts before a 14B render.

Best Wan 2.2 LoRAs for Creative Work

Because Wan’s encoder never refuses, LoRAs are the only thing between you and a new style or subject. These are the most useful ones as of November 2026:

Rules for stacking: put speed LoRAs first, then motion fixes, then style LoRAs, and keep each style LoRA at 0.6 to 1.0. LoRAs trained for Wan 2.1 often work on the low-noise expert of 2.2 but not on the high-noise expert, so test both. Always load a high-noise and a low-noise version when the LoRA comes in a pair.

Prefer an all-in-one file? Merged “Rapid AllInOne” checkpoints bundle the 4-step LoRAs into one GGUF, but you lose control over strengths and versions, so we recommend the separate files above.

Fun Camera, VACE and Animate

• Fun Camera Control (wan2.2_fun_camera_high_noise_14B_fp8_scaled.safetensors + low-noise): set an exact camera path (pan, tilt, zoom, orbit) instead of describing it. Template: Wan2.2 Fun Camera.
• Fun Control (wan2.2_fun_control_*): drive motion with pose, depth or canny video.
• Fun VACE (wan2.2_fun_vace_*): edit, extend or inpaint existing video.
• Animate 14B (wan2.2_animate_14B_int8_convrot.safetensors): animate one of your original characters with motion from a reference video. Wan-AI also published an Animate 2 checkpoint in August 2026 (Diffusers format, Apache 2.0); ComfyUI support was not documented at the time of writing.

All of these go in ComfyUI/models/diffusion_models/ and work with the same text encoder, VAE and 4-step LoRAs.

Prompting Wan 2.2

• Structure: subject, action, setting, camera, lighting, style. Example: “A woman in a red raincoat runs across a neon-lit street at night, rain splashing, handheld tracking shot, cinematic, shallow depth of field.”
• Describe motion explicitly. Wan follows verbs well: “turns her head and smiles”, “camera slowly orbits left”.
• Keep it to one action per 5-second clip. Chain clips for longer scenes.
• Negative prompt: the template’s default (blurry, static, extra limbs and so on) is good. At CFG 1, the negative prompt has little effect.
• Use a local LLM to expand prompts if you want longer descriptions. Any model from our uncensored local LLM list works, and it keeps your prompts private.

Troubleshooting

• Out of memory on 16 GB: drop to Q4_K_S or Q3_K_M, lower resolution to 832×480, or reduce length to 49 frames. Launching with --reserve-vram 1 can help (see ComfyUI launch flags).
• Video looks like slow motion: add the Slow-Mo Fix LoRA pair, or try the 1022 LightX2V I2V version.
• Washed-out or burned colours: CFG is too high for the 4-step LoRAs. Set CFG 1.0.
• Grey or noisy output from the 5B model: you loaded the 2.1 VAE. The 5B needs wan2.2_vae.safetensors.
• “Unet Loader (GGUF)” missing: ComfyUI-GGUF is not installed or failed to import. Check the console.
• More: ComfyUI errors and fixes.

FAQ

Can Wan 2.2 run on 16 GB VRAM?

Yes. Use the 14B GGUF Q4_K_M experts (9.65 GB each) with the LightX2V 4-step LoRAs at 832×480. Only one expert is in VRAM at a time.

Can Wan 2.2 run on 8 GB VRAM?

Yes, with the TI2V 5B model or the 14B at Q3 with offloading, and at least 32 GB of system RAM.

Is Wan 2.2 censored?

No. There is no filter in ComfyUI and its umT5 text encoder has no refusal training. What it cannot draw well comes from its training data, which LoRAs can extend.

Which LightX2V LoRA version should I use?

For I2V, the 1022 version; for T2V, the 1217 version, both from lightx2v/Wan2.2-Distill-Loras. The Comfy-Org v1 and v1.1 copies also work.

Where do Wan 2.2 LoRAs go?

ComfyUI/models/loras/. Load them with LoraLoaderModelOnly, one per expert.

Can I use Wan 2.2 commercially?

Yes. Wan 2.2 is Apache 2.0. Check the licence of each LoRA separately.

Final Thoughts

Wan 2.2 remains the most practical open video model in 2026: no refusals, a permissive licence and the biggest LoRA library. On a 16 GB card, the winning setup is the 14B I2V at GGUF Q4_K_M, the LightX2V 4-step LoRAs at CFG 1, and the Slow-Mo Fix when motion drags. Start from a strong still image, describe motion clearly, and add style LoRAs for the look you want.

Related guides:

• Uncensored Local AI Video Generation in ComfyUI
• Best Local AI Video Models in 2026
• ComfyUI Model Folder Cheat Sheet
• GGUF Quantization Explained
• 1080p 60 FPS With RIFE and SeedVR2

Sources

• ComfyUI docs: Wan 2.2 native workflows
• Comfy-Org/Wan-2.2 file list
• QuantStack Wan 2.2 I2V GGUF and T2V GGUF
• lightx2v/Wan2.2-Distill-Loras
• Wan 2.2 I2V LightX2V Slow-Mo Fix
• Comfy blog: Wan 2.2 Fun and LightX2V 4-step support
• ComfyUI Wiki: Wan 2.2 workflows
• Hugging Face model search for “wan2.2”, sorted by downloads (checked November 2, 2026)