Open-Source AI Video Creation: Wan 2.2 + LTX-2 + ComfyUI (Local Setup)

Essential Nodes to Install:

• ComfyUI-WanVideoWrapper (for Wan 2.2)
• ComfyUI-LTXVideo (for LTX-2)
• RTX Video upscaler node
• ComfyUI-VideoHelperSuite (for video I/O)

Step 2: Download Model Weights (1 hour, depending on internet)

Wan 2.2 Models (pick based on VRAM):

• Wan2.2-T2V-A14B (~28GB) — text-to-video with MoE
• Wan2.2-I2V-A14B (~28GB) — image-to-video
• Wan2.2-Animate-14B (~28GB) — character animation
• Wan2.2-T2V-1.3B (~3GB) — lightweight, runs on 8GB VRAM

LTX-2 Models:

• LTX-2 NVFP8 (~12GB) — quantized for RTX GPUs, best speed/quality
• LTX-2 BF16 (~38GB) — full precision, maximum quality

All models available on Hugging Face. Download with huggingface-cli.

Step 3: Text-to-Video with Wan 2.2 (Creative Phase)

Workflow: Rapid Exploration

Use Wan 2.2 T2V-A14B for initial creative exploration:

1. Write your prompt with cinematic details
2. Generate at 480p for fast iteration (5-10 seconds per clip)
3. Refine prompts based on results
4. Generate final at 720p

Prompt Template:

[camera movement], [subject description], [action], [environment], [lighting], [style]

Example: "Slow dolly forward, a woman reading in a sunlit library, turning pages, warm natural light through tall windows, cinematic film grain"

Wan 2.2 Style Control: The T2V-A14B model supports aesthetic data curation for cinematic style. Add style keywords:

• "cinematic, 35mm film" — Hollywood look
• "anime, vibrant colors" — Animation style
• "documentary, natural lighting" — Realistic look

Step 4: Upscale with LTX-2 (Quality Phase)

4K Upscaling Pipeline:

1. Take your best Wan 2.2 720p output
2. Feed it into LTX-2's upscaler via ComfyUI
3. LTX-2 upscales to 4K while adding detail (not just stretching pixels)
4. RTX Video node handles real-time sharpening and artifact cleanup

Adding Audio with LTX-2: LTX-2 is the first open-source model with joint audio+video generation:

1. Use LTX-2 S2V mode with your video as input
2. Provide an audio description prompt
3. LTX-2 generates synchronized audio matching the visual content

Step 5: Post-Processing with FFmpeg

Final Assembly:

1. Combine upscaled video with generated audio
2. Apply color correction (optional)
3. Export in target format and resolution

FFmpeg Command Template:

ffmpeg -i video.mp4 -i audio.wav -c:v libx264 -crf 18 -c:a aac -b:a 192k output.mp4

Cost Analysis: Local vs Cloud

Break-even point: If you generate 100+ videos/month, local pays for itself within a year.

Tips for Success

1. Start with Wan 2.2 1.3B to test your setup before downloading large models
2. Use weight streaming if you hit VRAM limits — it's slower but works
3. Save your ComfyUI workflows as JSON — share and reuse them
4. Join the ComfyUI Discord — the community is incredibly helpful
5. Monitor GPU temperature — sustained generation can heat up your system