Most video workflows stop at 10–15 seconds. H3MD Director is a free, open-source ComfyUI node pack that renders 2–3 minute videos with MiniMax H3 in one run – one continuous take, or a sequence of planned shots.
中文简介在下方 ↓
At a glance
• Two modes: 0.3.x renders segment by segment – each segment continues from the last 1.6 s of the one before. 0.4 adds Seamless long video (CCW, Chained Context Windows) – the whole take is rendered on one timeline. One tick box on the Director switches between them.
• Measured on one RTX 5090: a 2 min 20 s 1080p music video with Seamless long video in 2 h 18 min (10 shots, 3 groups, 48 fps), compute VRAM peak 11.1 GB. Full numbers in "About this version" on the right.
• VRAM: 16 GB minimum, 24 GB+ recommended; the 4B text encoder option saves 10+ GB.
• What you can make: music videos with lip-sync, motion copy, text-to-video. On the Director panel you plan every shot – length, prompt, reference pictures, locked audio.
• Examples: the gallery above; open the showcase post to see each clip's settings.
What's new in 0.4
CCW – Chained Context Windows: tick "Seamless long video" on the Director. Continued shots render together on one timeline – no continuation seams, no drift, no double images where windows overlap. Hard cuts stay clean. Needs ComfyUI 0.44+.
Redraw a few seconds: the H3MD Redraw / Preview node re-renders only the seconds you type (e.g. 1:12-1:15); you can change that shot's prompt or pictures first. Everything else stays frame-identical.
8-bit cache in the example workflows: about 9 GB per minute of 1080p instead of 18 GB, no visible difference.
Frame interpolation (48 / 60 fps) no longer blends two shots at a hard cut. 0.4.3: lighter Control Panel – each LoRA dropdown loads the full list only when you open it.
Earlier, in 0.3: the simple workflow (one Control Panel + the Director), the "Which segments" box to redo only some segments, plan export / import as one zip, and optional drift correction on save.
Install
ComfyUI 0.43+ (native MiniMax H3); 0.44+ for Seamless long video.
ComfyUI Manager → search "H3MD Director", or clone the GitHub repo into custom_nodes.
Drag H3MD Director simple.json from the zip onto the canvas.
GitHub (docs, model table, prompt guide): https://github.com/1113800942/comfy-h3md-director
Tutorials on YouTube: https://www.youtube.com/channel/UC1H4UDuSqD6M7BqLUlVy9lA
Install + first long video (EP02): https://youtu.be/zvPqr6ihZic
Seamless long video + redraw a few seconds (EP03): https://youtu.be/k0jZIbaH-aA
Previous version page (v0.1.1): https://civarchive.com/models/3032684
Full technical details (measured speed and memory, model files and where they go, optional parts): open "About this version" on the right, or see the GitHub README.
Run it online (no GPU needed)
RunningHub – cloud ComfyUI with ready-made workflows.
Three ready-to-run H3MD apps on RunningHub (referral links: the invite code gets you a discount, and it also supports this project):
• AI short drama, 6 shots, swap the images and the story: https://www.runninghub.ai/ai-detail/2120401708354838530?inviteCode=aolwbd1n
• Long music video, one photo + one song: https://www.runninghub.ai/ai-detail/2120407904939876354?inviteCode=aolwbd1n
• Motion copy, your photo follows any video: https://www.runninghub.ai/ai-detail/2120413562564083713?inviteCode=aolwbd1n
H3MD Director workflow on RunningHub: https://www.runninghub.ai/zh-cn/post/2117897074309345281
中文简介
H3MD Director 是免费开源的 ComfyUI 节点包，用 MiniMax H3 一次生成 2–3 分钟的长视频：可以是一整个长镜头，也可以是按计划排好的一组镜头。
• 两种生成方式：0.3.x 是分段式，一段接一段生成，每段从上一段最后 1.6 秒接着往下走；0.4 增加了无接缝的窗口式（接续窗口 CCW），整条长镜头放在一条时间线上一起生成。导演台上勾一下就能切换。
• 0.4 无接缝长视频（接续窗口 CCW）：在导演台顶部勾「Seamless long video」，接续的镜头放在同一条时间线上一起生成，段与段之间没有接缝、不漂移，窗口重叠的地方也不重影；硬切照样干净。需要 ComfyUI 0.44 以上。
• 局部重绘：「H3MD Redraw / Preview」节点只重跑你填的那几秒（如 1:12-1:15），可以先改这个镜头的提示词或参考图，其余画面逐帧不变。
• 缓存改 8 位：1080p 每分钟约 9 GB（原来约 18 GB），画质看不出差别。
• 实测 RTX 5090：2 分 20 秒的 MV（1920×1088，10 个镜头，3 组，48 fps）2 小时 18 分钟，计算显存峰值 11.1 GB；比分段式慢约 25%。
• 安装：ComfyUI Manager 搜「H3MD Director」，或把 GitHub 仓库克隆到 custom_nodes，再把压缩包里的工作流拖进画布。国内镜像：huggingface.co/yj0942/H3MD-Director（打不开可以把 huggingface.co 换成 hf-mirror.com）。
• 教程视频是中文讲解（带中英字幕），第三集：https://youtu.be/k0jZIbaH-aA
• 没有显卡也能跑：RunningHub 上三个现成应用（邀请链接，用邀请码有优惠）：
– AI 短剧（6 镜头可换图换剧情）：https://www.runninghub.ai/ai-detail/2120401708354838530?inviteCode=aolwbd1n
– MV（一张照片 + 一首歌）：https://www.runninghub.ai/ai-detail/2120407904939876354?inviteCode=aolwbd1n
– 动作复刻（照片跟着视频动）：https://www.runninghub.ai/ai-detail/2120413562564083713?inviteCode=aolwbd1n
License
• H3MD nodes: AGPL-3.0. MiniMax H3 weights are not included and have their own license.
• The RunningHub links are invite links (new sign-ups get a discount).
Sample videos are our own, made with H3MD: v0.4.3 – Peter, from the 0.4 seamless music-video test; v0.3.3 – Eve, "Above the Clouds" (song made with ACE-Step 1.5).
Built on ideas and work from Luna, Niko, Marker, LBH-123-AI, Kijai, T8 and others – thank you. License: AGPL-3.0 with attribution.
Description
CCW – Chained Context Windows: seamless long video. Tick "Seamless long video" on the Director: shots that continue each other are rendered together on one timeline, so there are no continuation seams and no drift from segment to segment. Hard cuts stay clean cuts.
ComfyUI's core Context Windows average the overlap between windows; where the picture changes a lot, two different pictures averaged become double images (see the showcase: a see-through railing across his hands). CCW runs the windows in time order and each window takes over the overlap from the one before it – nothing is averaged.
New node H3MD Redraw / Preview: low-res preview only, or redraw a few seconds of a finished video (e.g. 1:12-1:15); you can change that shot's prompt or reference pictures first. Everything outside the range stays frame-identical.
Frame interpolation (48 / 60 fps) no longer blends two shots at a hard cut.
The example workflows cache frames as 8-bit: about 9 GB per minute of 1080p video instead of 18 GB, no visible difference.
Needs ComfyUI 0.44 or newer for Seamless long video (on older versions it says so clearly). Untick it and you get the 0.3.x segment-by-segment mode.
0.4.3: lighter Control Panel – each LoRA dropdown loads the full list only when you open it.
Measured on an RTX 5090 32 GB, 64 GB RAM: a 2 min 20 s music video at 1920×1088, 10 shots in 3 groups, 48 fps: 2 h 18 min, compute VRAM peak 11.1 GB. Redrawing its last shot (18.6 s, with a changed prompt and an extra outfit picture): 25 min 51 s. About 25 % slower than rendering segment by segment.
Tutorial (EP03): https://youtu.be/k0jZIbaH-aA
TECHNICAL DETAILS
Measured speed and memory
One machine: RTX 5090 32 GB, Ryzen 9 9950X3D, 64 GB RAM, ComfyUI 0.43 (0.44 for Seamless long video), PyTorch 2.13 + CUDA 13.0, hybrid b25-49 int8 model, 32B int8 text encoder, 4-step turbo LoRAs, 6 low-res + 3 high-res steps
• 0.4 Seamless long video, 1920×1088: a 2 min 20 s music video (10 shots, 3 groups, 48 fps) in 2 h 18 min, compute VRAM peak 11.1 GB; redrawing its last shot (18.6 s, with a changed prompt and an extra outfit picture) took 25 min 51 s. About 25 % slower than rendering segment by segment. 8-bit cache: 6.3 MB per 1920×1088 frame (about 9 GB per minute).
• Segment by segment, 704×1280: about 4 min per ~15 s segment; a 2 min 40 s video (12 segments) in 47 min; 1–2 high-res strips; peak system RAM about 54 GB; disk cache about 22 GB.
• Segment by segment, 1088×1920 / 1920×1088: about 8–12 min per ~15 s segment; a 2 min 40 s video (11 segments) in about 1.5–2 h; 2–3 high-res strips; peak system RAM about 35 GB; disk cache about 70 GB.
• The first segment of a run takes about 1 minute longer (model loading).
• 4B text encoder + ClipProj instead of 32B: same picture quality and speed (203 s vs 199 s per segment at 704×1280, same seed), saves 10+ GB VRAM, follows long prompts slightly less closely.
• Cache: about 3 GB per 243 frames at 1088×1920; delete it after saving (delete_cache_after_save).
• VRAM: the high-res pass is split into strips automatically from the free VRAM, so the peak follows your card. Only tested on an RTX 5090 so far. Minimum 16 GB, 24 GB+ recommended; smaller cards start with 4B + ClipProj, 704×1280 and ~10 s segments.
Models (file → ComfyUI models folder)
• minimax_h3_hybrid_fl2va_ref2va_b25-49-int8.safetensors → diffusion_models (huggingface.co/smhfacct/Minimax-H3-fl2va-ref2va-hybrid-models)
• Text encoder → text_encoders: 32B qwen3vl_32b_minimax_h3_int8_convrot, or 4B Huihui-Qwen3-VL-4B-Instruct-abliterated-int8_convrot.safetensors (huggingface.co/ahmed22xa/Huihui-Qwen3-VL-4B-Instruct-abliterated-comfy)
• With the 4B encoder only: mmh3-4b-ClipProj-v3.1-mlp.safetensors → clip_projections (huggingface.co/NicoLab28/ClipProj-MiniMax-H3)
• minimax_h3_video_vae_fp16.safetensors and minimax_h3_audio_vae_fp32.safetensors → vae (huggingface.co/Comfy-Org/MiniMax-H3)
• h3_upscaler_sharpness_2000steps_v0.1_fp32.safetensors → latent_upscale_models (huggingface.co/Alissonerdx/Minimax-H3-ComfyUI)
• LoRAs → loras: minimax_h3_fl2v_turbo_4step_v1.2_768p_comfyui_bf16 (lightx2v/Minimax-h3-Turbo), minimax_h3_DMAD_4step_full_lora_avg_rank_39_bf16 (Kijai/MiniMax-H3-experimental), minimax_h3_lms_v1.0_r64_r (RunningHubAI/rh-minimax-h3-lms-v1.0-r64-lora), H3_Combat_V2 (JOKER141/MiniMax-H3-Combat-Base-V2)
• Sampling: two passes, 6 steps at low resolution, learned latent upscaler, 3 steps at full resolution (4-step turbo LoRAs).
• Sizes: base 19.5 GB, 32B encoder 25.3 GB (or 4B 5.3 GB + ClipProj 0.5 GB), video VAE 4.9 GB, audio VAE 0.6 GB, each LoRA 0.3–1.8 GB. Plan for about 55 GB of models plus the cache. The workflow's "Model links" note lists every file with its download link.
Optional parts (off by default)
• Face refine: SeedVR2 (numz) or VOSR2 node pack and its model. Use it only for widescreen shots where the face is small (¼ of the frame or less). Extra time and VRAM: not measured yet; numbers will be added here.
• Motion copy: the third-party TARA LoRA (Alissonerdx) + ComfyUI-BFSNodes – install them yourself; the Control Panel's "Motion copy" button switches them on. The older depth control (Fun ControlNet Union, Depth Anything V2, SAM3) is still available as an option.
• RIFE interpolation on save: ComfyUI-VFI with RIFE v4.26 flownet.pkl.