[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"project-94392":3},{"id":4,"name":5,"fullName":6,"owner":7,"repo":5,"description":8,"homepage":8,"htmlUrl":8,"language":9,"languages":8,"totalLinesOfCode":8,"stars":10,"forks":11,"watchers":12,"openIssues":13,"contributorsCount":14,"subscribersCount":14,"size":14,"stars1d":14,"stars7d":14,"stars30d":15,"stars90d":14,"forks30d":14,"starsTrendScore":14,"compositeScore":16,"rankGlobal":8,"rankLanguage":8,"license":17,"archived":18,"fork":18,"defaultBranch":19,"hasWiki":20,"hasPages":18,"topics":21,"createdAt":8,"pushedAt":8,"updatedAt":22,"readmeContent":23,"aiSummary":24,"trendingCount":14,"starSnapshotCount":14,"syncStatus":13,"lastSyncTime":25,"discoverSource":26},94392,"ComfyUI-MiniMax-H3-Turbo","Larryvrh\u002FComfyUI-MiniMax-H3-Turbo","Larryvrh",null,"Python",402,26,1,2,0,292,51.29,"Apache License 2.0",false,"main",true,[],"2026-08-24 04:01:22","# ComfyUI-MiniMax-H3-Turbo\n\nRun [MiniMax-H3](https:\u002F\u002Fdocs.comfy.org\u002Ftutorials\u002Fvideo\u002Fminimax\u002Fminimax-h3) —\njoint **video + synchronized audio** — in as few as **4 sampling steps** instead\nof ~20, with the\n[MiniMax-H3 Turbo LoRA](https:\u002F\u002Fhuggingface.co\u002Flarryvrh\u002FMiniMax-H3-Turbo-Lora).\n\nTwo nodes drop straight into the official H3 workflow (text-to-video and\nimage-to-video):\n\n| node | what it does |\n|---|---|\n| **MiniMax-H3 Turbo LoRA** | `MODEL → MODEL`, applies the turbo LoRA |\n| **MiniMax-H3 Turbo Sampler** | `→ SAMPLER`, feeds `SamplerCustomAdvanced` |\n\n## Which checkpoint — `v4` (600) or `v1` (850)?\n\nFor **most** work, use **`minimax_h3_turbo_v4_step600_ema.safetensors`** (from the\n[LoRA repo](https:\u002F\u002Fhuggingface.co\u002Flarryvrh\u002FMiniMax-H3-Turbo-Lora)). It's the\nstrongest checkpoint so far: much better static \u002F small-motion shots, markedly\nbetter micro-detail (faces, fingers, texture), and the over-sharpening \u002F plastic\nlook of the earlier `v1` (~850) line is fully resolved.\n\nv4 introduced a **static-frame enhancement**. The one trade-off shows up **only at\n4 steps with large, fast motion**, where v4 can produce **motion-smear \u002F trailing\nghosting** (actively being fixed). **Using 6–8 steps largely removes it** — and v4\ntolerates higher step counts better than v1 (which over-sharpens at high steps +\nstrength 1.0). For the narrow case of **4 steps *and* heavy motion**, the older\n**`v1` ~850** checkpoint can still be friendlier.\n\n```\nUsing 6–8 steps?        ── yes ──►  v4-600  (recommended)\n   │ no (4 steps)\n   ▼\nHeavy \u002F fast motion?    ── no  ──►  v4-600  (recommended)\n   │ yes\n   ▼\n                                    v1-850  (friendlier at 4-step heavy motion)\n```\n\nStill a preview — the two areas still being improved are **audio** and **fast,\nintense motion**.\n\n## Install\n\n**Via ComfyUI-Manager** — search \"MiniMax-H3 Turbo\" and install. **Or manually:**\n\n```bash\ncd ComfyUI\u002Fcustom_nodes\ngit clone https:\u002F\u002Fgithub.com\u002Flarryvrh\u002FComfyUI-MiniMax-H3-Turbo\n```\n\nThen restart ComfyUI. Keep the node updated (Manager, or `git pull`) — it evolves\nalongside the weights.\n\nPut the LoRA `.safetensors` into `ComfyUI\u002Fmodels\u002Floras\u002F`. You also need the base\nMiniMax-H3 model, VAEs and text encoder from the official release — see the\n[MiniMax-H3 tutorial](https:\u002F\u002Fdocs.comfy.org\u002Ftutorials\u002Fvideo\u002Fminimax\u002Fminimax-h3).\n\n## Use\n\nStart from the **official MiniMax-H3 workflow** (t2v or i2v) and make two changes:\n\n1. Insert **MiniMax-H3 Turbo LoRA** between the model loader and the sampler\n   (`… → Load Diffusion Model → MiniMax-H3 Turbo LoRA → SamplerCustomAdvanced`),\n   and pick the turbo `.safetensors`.\n2. Feed `SamplerCustomAdvanced` from **MiniMax-H3 Turbo Sampler**, and set the\n   scheduler node (`BasicScheduler`) to `simple` at **≥ 4 steps**.\n\nEverything else — conditioning, VAE decode, audio output — stays as in the\nofficial graph, so **both t2v and i2v** work unchanged. A ready-made t2v workflow\nis in [`example_workflows\u002F`](example_workflows\u002Fminimax_h3_t2v_turbo.json) — drag it\ninto ComfyUI to see the wiring.\n\n## Steps and strength\n\n- **4 steps is the recommended *minimum*; 4–8 is the useful range.** 6–8 steps\n  look noticeably better than 4, so add steps if you can afford them. Past **8\n  steps** it stops helping and can start to introduce **over-sharp artifacts** —\n  no benefit to going higher, so stay in **4–8**.\n- **Keep `strength` at `1.0`.** It's tuned for 1.0 and holds up across the 4–8 step\n  range. Only reach for the dial if a *specific* clip misbehaves: **blurry ghosting\n  \u002F smear → nudge up** (`~1.05–1.2`), **over-sharp grain → nudge down**\n  (`~0.8–0.95`).\n- Keep the scheduler on `simple`.\n\n## Base model & `low_vram`\n\n**Base model** — works with any MiniMax-H3 base: full (`bf16`, `int8_convrot`)\n**and the pruned\u002Fcurve variants** (`pruned_int8`, `pruned_fp8`). The node detects a\npruned base automatically and re-injects the LoRA's time-conditioning at run time\n(a small `silu(t_emb)` grid ships with the node for this), so **one LoRA file\ncovers every base**.\n\n**`low_vram`** (node switch) trades sharpness for peak VRAM:\n\n- **off (default)** — applies the LoRA at run time (bypass): sharpest, recommended,\n  a little extra peak VRAM.\n- **on** — merges the LoRA into the weights: lowest peak VRAM, so smaller GPUs run\n  and longer \u002F higher-res clips fit, but the result is **softer on quantized\n  (`int8` \u002F `fp8` \u002F pruned) bases** (the tiny update is partly rounded away when\n  folded into the quantized weights).\n\nThe node streams the base model, so it runs on much smaller GPUs than the ~33 B\nsize suggests — an 80 GB GPU is only needed for the largest resolutions in `bypass`\nmode. If you hit out-of-memory, turn `low_vram` **on** and\u002For lower the resolution\nor frame count.\n\n## Why a custom sampler (and how it adapts)\n\nMiniMax-H3 denoises the video and audio streams on two different flow schedules\n(video shift 12, audio shift 3). **Recent ComfyUI handles this natively** — its\n`ModelSamplingAV` carries the audio latent on the video schedule — so a stock\nsampler already produces correct audio there. On **older ComfyUI without that\nsupport**, a stock sampler steps both streams on one schedule and badly over-steps\nthe audio at 4 steps, so the audio comes out distorted.\n\nThis node's sampler **auto-detects which ComfyUI it's on**: on recent builds it\nsteps as a plain single-schedule sampler (bit-for-bit the stock result); on older\nbuilds it steps each stream on its own clock so audio stays clean at 4 steps. Keep\nit in the workflow and it does the right thing across ComfyUI versions — nothing to\nchange when you update. (On recent ComfyUI a stock `euler` also works; the Turbo\nSampler just keeps existing graphs running unchanged.)\n\n## Notes\n\n- **Resolution \u002F length**: width and height are multiples of 32 (short edge\n  typically 768); frame count is at 24 fps and snaps to the model's 17·k+5 grid\n  (124 ≈ 5 s). Validated range ~124–362 frames.\n\n## License\n\nApache-2.0.\n","这是一个为 ComfyUI 设计的轻量级加速插件，用于显著提升 MiniMax-H3 视频生成模型的推理效率。核心功能包括：通过 Turbo LoRA 微调模型、集成专用 Turbo Sampler（支持仅需4步采样的快速视频合成），并保持与原始 MiniMax-H3 文本\u002F图像到视频工作流完全兼容；技术上依赖 LoRA 微调、自定义采样调度（simple scheduler）及静态帧增强优化。适用于对生成速度敏感的短视频创作、AIGC 快速原型验证、以及需在有限算力下运行高质量音画同步视频生成的本地部署场景。","2026-08-07 02:30:09","CREATED_QUERY"]