Got VACE working in real-time
30fps on a 5090
Details
- External ID
- 46989470
- Source
- HN
- Company
- —
- Product
- Got VACE working in real-time
- Website domain
- daydream.live
- Launched
- Feb. 12, 2026
- Cohort
- —
- Upvotes
- 10
- Upvotes percentile
- 0.5316711590296496
- Tags
- —
- Fetched at
- Sept. 7, 2026, 9:25 p.m.
- Updated at
- Sept. 7, 2026, 9:25 p.m.
Description
I adapted VACE to work with real-time autoregressive video generation.Here's what it can do right now in real time:- Depth, pose, optical flow, scribble, edge maps — all the v2v control stuff - First frame animation / last frame lead-in / keyframe interpolation - Inpainting with static or dynamic masks - Stacking stuff together (e.g. depth + LoRA, inpainting + reference images) - Reference-to-video is in there too but honestly quality isn't great yet compared to batchGetting ~20 fps for most control modes on a 5090 at 368x640 with the 1.3B models. Image-to-video hits ~28 fps. Works with 14b models as well, but doesnt fit on 5090 with VACE.This is all part of Daydream Scope (https://github.com/daydreamlive/scope), which is an open source tool for running real-time interactive video generation pipelines. The demo was created in scope, and is a combination of Longlive, VACE+Scribble, Custom LoRA.There's also a very early WIP ComfyUI node pack wrapping scope: https://github.com/daydreamlive/ComfyUI-Daydream-ScopeCurious what people think.
Enrichment
- Theme
- AI video generation and ComfyUI tools
- Vertical
- Horizontal
- Function
- Hardware & robotics
- Audience
- Developer
- AI stance
- Not AI
- Project type
- Hobby / open-source project
- Normalized one-liner
- real-time video processing on gpu
- Manually corrected
- False
Could you build this?
No Achieving 30fps real-time autoregressive video generation and video-to-video conditioning on consumer hardware requires custom CUDA kernels, tensor optimization, and deep generative video modeling expertise.
What it would actually take: A real version requires adapting diffusion/autoregressive video architectures (like VACE) with TensorRT, FlashAttention, and custom CUDA C++ kernels to maximize GPU memory bandwidth on RTX 50-series cards. Techniques like speculative decoding, latent cache reuse, and pipelined frame conditioning are necessary. This demands senior GPU systems researchers with deep knowledge of PyTorch internals and high-performance graphics compute.
Discussion
No comments on this launch.
Competitors
Other products that read as similar to this one — 108 launches clear the similarity bar, closest 8 shown.
Attention rank: #50 of 109 (itself plus its competitors, highest first — normalized so YC and Product Hunt are compared fairly).
Launched 103 days after the earliest competitor.
- Virtual time engine to record HTML to 60fps MP4, no drops · hn · 2026-07-26 · 5 upvotes · similarity 0.47
- ComfyUI-Qwen-VAE-Triton · github · 2026-09-13 · 7 upvotes · similarity 0.45
- lingbot-world-v2-realtime · github · 2026-09-16 · 14 upvotes · similarity 0.42
- Crow-DLSS5-Video-Image-Converte · github · 2026-09-11 · 7 upvotes · similarity 0.41
- FastVSR · ph · 2026-09-15 · 1 upvotes · similarity 0.40
- LemonSlice · hn · 2026-01-27 · 133 upvotes · similarity 0.40
- FrameSketch · ph · 2026-09-12 · 67 upvotes · similarity 0.40
- Text-to-video model from scratch (2 brothers, 2 years, 2B params) · hn · 2026-01-22 · 158 upvotes · similarity 0.40
Other launches for this product
- No other launches for this product.
Same idea, different domain
Nobody's really built a hardware & robotics tool for Fintech yet.