- Home
- Skills
- Image & Video
- Wan 2.2 Image-to-Video, Tuned for ComfyUI
More screenshots
Works with the AI Agents you already use
Wan 2.2 Image-to-Video, Tuned for ComfyUI
Two tested Wan 2.2 image-to-video workflows for ComfyUI on 24 GB GPUs: 6.4 s clips in about 4 minutes, plus the decode, clip-length and prompting rules that make them work
$3.99
Wan 2.2 Image-to-Video, Tuned for ComfyUI
Example session with this skill installed
I need to generate a 5-second cinematic portrait video using Wan 2.2 TI2V-5B. The render keeps running out of memory during the VAE decode. Help me fix the settings.
- Read your context and instructions
- Compiled the wan 2 2
Fixed — your 5B decode was asking for one 18.6 GB block of VRAM.
At 704×1280 × 121 frames, a plain VAEDecode decodes the whole clip in one piece and runs out of memory after ~11 minutes of sampling. The fix is to tile
in space only, never in time:
| Setting | Value | Why |
|---|---|---|
| Node | VAEDecodeTiled | replaces VAEDecode |
| tile_size / overlap | 512 / 64 | splits each frame into tiles |
| temporal_size | 256 | larger than the clip, so it is never cut in time |
| temporal_overlap | 8 | unused at this length |
Tiling in time breaks Wan's causal VAE cache, so the picture drifts more the longer the clip runs (43 dB → 31 dB PSNR by frame 80). Spatial tiles don't have that problem.
Measured result (RX 7900 XTX 24 GB): decode 56 s, no visible tile seams, 121 frames = 5.04 s @ 24 fps, no rewind to the first frame.
Also keep the VRAM_Debug (unload_all_models) node between the sampler and the decode: it frees the ~12 GB model before decoding.
Next steps
- Load
workflows/Wan22_5B_I2V.json— the fix is already in it. - Use a portrait start image (square images get stretched).
- Want a longer clip? See the length rules — use RIFE, not more frames.
Connects securely to your tools. The creator never sees your data.
About this skill
High-fidelity video generation often stalls at the VAE decode stage or suffers from temporal artifacts like rewinding and flickering. This skill provides two verified ComfyUI workflows and a set of precise configuration rules for Wan 2.2 Image-to-Video models. It solves the "240-second decode" problem with targeted VRAM management and tunes both the A14B and 5B models for 24 GB GPUs.
What it does
- Fixes the slow decode by unloading the diffusion model right before the VAE decode, so the decode no longer spills into shared memory (244 s → 12.6 s, measured).
- Prevents temporal drift by never tiling the decode in time: the A14B workflow decodes in one piece, and the 5B workflow tiles in space only (
temporal_sizelarger than the clip). - Speeds up sampling with the 4-step Lightning LoRAs for A14B at CFG 1.0, euler/simple.
- Smooths motion with RIFE x2 interpolation: -46% flicker and -35% background wobble, no ghosting.
- Avoids broken lengths with tested frame limits: 97 frames is safe, 121 rewinds to the first frame, 161 hits a VRAM cliff.
How it works
- Load the workflows into ComfyUI and install
ComfyUI-GGUFandComfyUI-KJNodesvia the Manager. - Pick a workflow: A14B (GGUF Q6_K) for the best quality per second, or TI2V-5B (fp16) for 704×1280 portrait video.
- Write motion prompts: describe camera movement and action with time segments, not the scene. The start image already sets the scene.
- Run it with the included
VRAM_Debugpurge node left in place, since it is what keeps the decode fast.
Frameworks & tools
Built for ComfyUI with ComfyUI-GGUF and KJNodes, RIFE v4.26 and the umt5 text encoder. Tested on an RX 7900 XTX (24 GB) with ComfyUI 0.25 and 0.45.2. The memory and clip-length rules come from the model itself, so they should apply on NVIDIA too, but that was not tested. On AMD, the companion AMD skill adds SageAttention for about 1.65x faster sampling.
Why this beats prompting it yourself
Generic advice misses the VRAM bottleneck between sampling and decoding, which can make the VAE decode up to 20x slower than it needs to be (244 s instead of 12.6 s). This skill also captures Wan 2.2's measured frame limits and decode-tiling behaviour, which are not documented in the model cards.
Requires 24GB VRAM. A14B length capped at 97 frames to prevent rewinding. A14B negative prompts are inert at CFG 1.0. RIFE requires Comfy-Org specific build.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 4 days ago
Requires