Bij de wetsbehandeling: iedereen gaat erop vooruit in het nieuwe pensioenstelsel
En nu keurt de toezichthouder DNB een plan goed waardoor je er 70% op achteruit kunt gaan.
We waarschuwden en dit gaat leiden tot rechtszaken en meer.
pensioenpro.nl/jonge-slaper-…
🔥 Running WAN 2.2 on Only 2 GPUs — Real‑World Test
I’ve been testing WAN 2.2 on a 4×3090 rig, but I needed GPU 0 1 free for Qwen.
So this whole experiment was done using only **GPU 2 GPU 3**.
Here’s what actually works 👇
1️⃣ TI2V‑5B *does* run on 24GB cards — but only with CPU offload
The VAE is too big for a single 3090.
The trick is simple:
• Diffusion on GPU
• VAE T5 on CPU
• Use system RAM as the buffer
Zero CUDA OOM errors.
2️⃣ You can generate long videos (6–8 sec) reliably
Using:
--offload_model True
--t5_cpu
--ulysses_size 1
--infer_frames 32–64
It’s slow, but rock‑solid.
3️⃣ Best way to use 2 GPUs = run 2 jobs in parallel
Forget multi‑GPU mode for TI2V‑5B — the VAE doesn’t split.
Real‑world fastest method:
• GPU 2 → Job A
• GPU 3 → Job B
Two videos at a time, stable.
4️⃣ You can automate 20 clips overnight
I built a batch script that:
• Reads prompts from a file
• Runs 2 WAN jobs at a time
• Queues the rest
• Saves everything cleanly
• Never touches GPU 0 1
Perfect for building a full DnB music video from AI‑generated clips.
5️⃣ Qwen handles the orchestration
Qwen on GPU 0 1 can:
• Generate prompts
• Build the batch script
• Manage the queue
• Organize outputs
• Help assemble the final video
AI controlling AI.
6️⃣ Next step: full automated DnB video pipeline
20 clips → 1 music video.
All generated on GPU 2 3 while Qwen stays free on GPU 0 1.
This setup works — and it’s only the beginning.
#Wan2.2 #LocalAI#AIvideo#TextToVideo#OpenSourceAI#RTX3090
Frazer, 18, was shown a traumatic, violent video by another pupil at school when he was in Year 8, without his consent. He's since been diagnosed with PTSD and has suffered for years with his mental health.
🔗 trib.al/lW912lw