ComfyStreamerH3 runs MiniMax H3 video on one RTX 5090 at 448×256
Comfy reports 1.28× aggregate throughput in a warm low-resolution run; the published figures do not establish single-clip latency.
Comfy described ComfyStreamerH3, a ComfyUI node for MiniMax H3 video generation on one 32 GB RTX 5090, in an 8 October 2026 blog post. The experimental workflow targets 448×256-pixel output.
The project README lists two 15.08-second clips, a median 23.51 seconds across 11 warm RTX 5090 jobs and four-step VSA. Its stated 1.28× real-time figure uses the combined clip duration and measures aggregate throughput. It does not establish single-clip response time or cold-start latency.
Bryan Wade’s post describes combining FastH3 V2 at four sampling steps with sparse attention, a smaller text encoder, pruned and quantized weights, and overlapping decode and encoding. Wade acknowledges artifacts and remaining limits in detail, consistency and resolution. This is low-resolution experimental generation rather than delivery-ready footage.
The managed runtime targets Linux, an Nvidia GPU, Python 3.12 and CUDA 13 PyTorch. The node code is MIT-licensed, but the repository supplies no model weights and third-party components retain their terms. Model licences require separate checking. Motion Engineering Today has not reproduced the benchmark or assessed the output.
AI-assisted, reviewed by Codex (15-post review authorized by JJ Rooney, 15 October 2026) on 15 October 2026.
