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.
At a glance
• Measured on one RTX 5090: a 2 min 40 s music video at 1080p in about 1.5–2 h (11 segments), about 8–12 min per 15 s segment. Full numbers at the bottom.
• VRAM: 16 GB minimum, 24 GB+ recommended; the 4B text encoder option saves 10+ GB.
• Examples: the gallery above; open the showcase post to see each clip's settings.
What's new in 0.3
Simple workflow: one Control Panel + the Director on the canvas; models, sampler and save are tucked into subgraphs.
Control Panel: size, fps / interpolation, face refine, seed, style buttons (Balanced / High motion / Motion copy) and the low-res / high-res LoRA lists in one node – it loads the LoRAs itself.
"Which segments" box: type 1 to test only the first segment, 2, 5-7 to redo a few; clear it to render the whole video (finished segments come from the cache).
Save and share a plan: export the plan + media as one zip, import it on another machine.
Drift correction on save (optional): brightness, colour, saturation and sharpness measured per segment and pulled back when they drift.
Works with T8's semantic bridge and with the plain SeedVR2 (numz) or VOSR2 for face refine.
What it does
Long takes: each segment continues from the last 1.6 s of the previous one (latent overlap), so motion and camera carry on with no cut.
Hard cuts: a new shot can start fresh from its own reference images, with no accumulated drift.
Director panel: plan shots, lengths, prompts, reference images and locked audio per shot.
Music videos with lip-sync, motion transfer (depth + SAM3), text-to-video.
1080P for long videos; low-VRAM / tiled options included.
Install
ComfyUI 0.43+ (native MiniMax H3).
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
Previous version page (v0.1.1): https://civitai.com/models/3032684
Run it online (no GPU needed) RunningHub – cloud ComfyUI with ready-made workflows. H3MD Director app: https://www.runninghub.ai/ai-detail/2117891615172308993?inviteCode=aolwbd1n (Referral link: sign up through it and you get 500 RH coins free. It also supports this project.)
H3MD Director : https://www.runninghub.ai/zh-cn/post/2117897074309345281
Technical details (measured numbers, models, optional parts)
Measured speed and memory
One machine: RTX 5090 32 GB, Ryzen 9 9950X3D, 64 GB RAM, ComfyUI 0.43, 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
• 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.
• 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: alibaba-pai MiniMax-H3 Fun ControlNet Union, Depth Anything V2, SAM3. Extra time and VRAM: not measured yet; numbers will be added here.
• RIFE interpolation on save: ComfyUI-VFI with RIFE v4.26 flownet.pkl.
License
• H3MD nodes: AGPL-3.0. MiniMax H3 weights are not included and have their own license.
• The RunningHub link is an invite link (500 free RH coins for new sign-ups).
Sample video: Eve, "Above the Clouds" – our own music video (song made with ACE-Step 1.5), 1920×1088, made with H3MD.
Built on ideas and work from Luna, Niko, Marker, LBH-123-AI, Kijai, T8 and others – thank you. License: AGPL-3.0 with attribution.