MiniMax H3 单卡 RTX 5090 实时视频生成
Comfy 博客介绍了一个开源自定义节点 ComfyStreamerH3：在单张 32GB 显存的 RTX 5090 上运行 MiniMax H3，用四步采样在 15 秒内生成 15 秒、448×256 的视频。作者列出的优化包括稀疏注意力（只对约 20% 视频块做完整注意力）、用 NicoLab28 的 ClipProj 以 Qwen3-VL-4B 替代 H3 的 32B 文本编码器（约 4.5GB 对 15.7GB）、FastVideo 的剪枝 INT8 权重、融合 FP4 MLP，以及 Kijai 的 INT8 分块解码器与 NVENC 编码重叠，使显存需求从 80GB 降至 30GB 以下。文中称持续运行这类模型每天成本可高达 6000 美元，而 RTX 5090 租用最低约每小时 0.7 美元；作为对比，FastVideo Preview V1 在单张 B200 上生成 15 秒 1344×768 片段耗时 47.2 秒。作者承认画面仍有明显瑕疵，并建议后续用 SolRefinery 等上采样模型，把 448×256 的快速初稿低成本放大。
若此类优化继续累积，消费级单卡上的低成本准实时视频生成可能比预期更早到来。
ComfyStreamerH3
Background
Live video has been the talk of the town on AI twitter these days. From Fal’s post trained H3 Max model that competed on quality and topped the charts last month, to the FastH3 model that optimized on cost. Video models have come a long way since people joked about AI models being unable to animate Will Smith Eating Spaghetti! As impressive as these recent model are, there exists a problem. And that problem is cost. Running these models continuously can cost an upwards of 6k per day! This makes these models completely unusable for the average consumer.
Thanks for reading ComfyUI Newsletter! Subscribe for free to receive new posts and support my work.
The goal
So the question I asked is how far can these models be pushed? What if you had an agent search the internet for every optimization, every speed up, every cost savings, and definitely every dirty hack that sacrificed quality for speed. And then what if you put it all together into a single custom comfy node that runs on a single consumer GPU? My requirements were the following:
#MiniMax#MiniMax H3#ComfyUI#ComfyStreamerH3#FastVideo#FastH3 V2#Nvidia RTX 5090#Qwen3-VL-4B#SolRefinery#Nvidia B200