HR Endless Sampler: open-source ComfyUI node renders any-length H3 videos on 16GB VRAM
rhradec · reddit · 2026-08-30
Developer hradec released ComfyUI-HR-Endless-Sampler (alpha), removing Minimax H3's duration and VRAM limits: any-length video at any resolution on just 16GB of VRAM — the author rendered a full 600-frame 1080p video. It's open source with a template workflow.
How it works: the sampler node splits a long video into chunks, automatically attaching the last frames of the previous chunk as continuation. It also uses Gemma 3 12B QAT as a "chunk director", splitting the overall prompt into per-chunk prompts, watching previous chunks for continuity, and verifying each chunk matches the intended time-slice of the overall action.
Companion nodes: a live preview (built on KJNodes with TAEH3 tiny VAE, frame-by-frame walking, hover shows each chunk's Gemma prompt and render time), plus save/load nodes that persist per-chunk prompts, timeline and timings. The save node, from the author's VFX background, decodes latents internally and exports unclamped floating-point EXR.
The author notes low-VRAM users benefit most, but high-VRAM users can also break H3's 15-second barrier and even render 4K or 8K.
More from Infra
- Grok claims 816B MoE model runs locally on dual M5 Ultra Mac Studios via MLX — grok · 2026-10-16
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- DuoDiscover agents auto-build fast eval proxies, scoring 2.68x TPU kernel speedup — berkeley_ai · 2026-10-16
- Google Raises 2026 Capital Spending to Up to $205 Billion — emmanuelvivier · 2026-10-16