Tencent's Hy-Nomas video world model ported to native ComfyUI, runs on 12GB VRAM
LatentSpacer · reddit · 2026-10-24
- A Reddit user ported Tencent's Hy-Nomas video world model to native ComfyUI nodes: the DiT runs as a standard ComfyUI MODEL sampled by the stock KSampler with ComfyUI's memory management, needing only 12GB VRAM.
- Features: images are encoded as a still clip into the model's memory as the starting world; segments chain with per-segment prompts; camera control via key timelines or a 3D path recorder; a WASD drive mode generates 1-second turns (4.3s per turn at 832x480 on a 4090, not real-time).
- Performance: Tencent's INT8 weights via Triton block-scaled GEMM plus fused kernels, FP8 decoder and int8 memory give 1.02s per 720p step (vs 2.29s plain bf16); 10s of 832x480 takes 34s end-to-end, a 20s 720p world 3 minutes; 12GB vs 16GB speeds are nearly identical.
- Tip: write prompts in Tencent's prompt-rewriter format (<|Subject1|, <Action1>, follow-ups starting "Compared to the previous segment") or later segments get ignored.
- Nodes and weights are open-sourced on GitHub and Hugging Face; more checkpoints, LoRAs and a faster decoder are planned.
Related event: Tencent's Hy-Nomas Video World Model Runs Natively in ComfyUI(2 posts)→
More from coding & agent
- 8 Claude Code prompts that turn your spreadsheets and docs into charts, no design software needed — alex_verem · 2026-10-24
- Codex day 19: switching reasoning mid-conversation no longer invalidates cache, CLI updated — pvncher · 2026-10-24
- draft-cards: Hermes plugin turns AI email drafts into reviewable, approval-first cards — Teknium · 2026-10-24
- 217 lines of pure SQL render a pelican on a bike via SDF raymarching, no SVG or bitmaps — fofrAI · 2026-10-24
- OpenClaw Enterprise positions itself as the Kubernetes control plane for AI agent fleets — heyneighbor · 2026-10-24
- Factory tested Reflection's Beam early via Together AI on real coding and agentic workflows — togethercompute · 2026-10-24