Fix blurry distant faces in MiniMax H3 videos! 🎬✨ I built a 2-stage ComfyUI pipeline combining H3 Singularity + H3-FaceRefine + VOSR2 to surgically restore facial details: 1️⃣ YOLOv8-Face auto-tracks & crops 768x768 face canvas 2️⃣ Dynamic per-frame denoise ("fix small faces more, big faces less") 3️⃣ VOSR2 1x... super-detail refinement 4️⃣ Seamless feathered stitch-back Full breakdown & RunningHub workflows below: 👇 🔗 #MiniMaxH3 #ComfyUI #FaceRestoration #AIVideo #VOSR2 #GenerativeAIshow more

The per-frame denoise schedule is a nice idea. I’d smooth that schedule over time too: as a face crosses a size threshold, a sudden strength change could cause texture popping even with stable tracking. Have you compared smoothed vs independent frame settings?

Tryed that too — you'll see the cropped area's quality difference on realistic or anime content. Better to just do a 2nd pass on the whole video with a lower denoise; it'll also be more native, no additional custom nodes needed.

这套拆分很对：先裁脸再做动态去噪，确实比全帧硬修更省算力。真正的坑是 YOLO 框抖动带来的 crop jitter/identity drift——你是用 track smoothing 还是复用上一帧框？768² 下峰值 VRAM 也很想看。