Tired of blurry fingers, noise artifacts, and messy, tangled node graphs for MiniMax H3?
This workflow is built for high-performance, rock-solid stability, and long-form video generation.
Designed specifically to run efficiently on 16 GB VRAM GPUs (like the RTX 5060 Ti) while delivering professional-grade results.
UPDATE v1.1 Changes
Visuals
- Greatly simplified and cleaner layout for better readability
Fixes
- Fixed first pass metadata/info handling
- Fixed Single Pass mode (already patched in v1.0b)
- Fixed Supersample / Upscale toggle logic
- Resolved various minor bugs
Key Features
True Dual-Pass Efficiency: Starts with an ultra-fast 0.21MP structural pass for motion and physics, followed by a 1MP latent refinement pass for crystal-clear fine details.
Pixel-Perfect Scaling: Precise mathematical aspect-ratio alignment between passes to avoid any stretching or blur.
Clean Patch Order: Solves hand and face jitter by maintaining mathematical headroom and utilizing a clean attention patch sequence (
ModelAttentionBackend->ModelSampling->Chunk FF->EasyCache).Motion Locked: Integrated with
BetterMoveV2LoRA for seamless camera rotations and fast character movement.Clean & Modular: Fully annotated, neatly organized, and free of useless node clutter.
User-Friendly Control Panel
No need to dig into complex wire webs. Control your entire generation pipeline from a dedicated, clean control panel:
Dual-Pass & Upscale Toggles: Switch seamlessly between Latent Upscale and RTX/Supersampling upscalers depending on your GPU resources.
First Pass Ratio Control: Easily fine-tune your structural pass ratio (0.21 by default for optimal VRAM/speed balance).
NVIDIA RTX Post-Processing Suite: Built-in toggles for RTX Upscale, RTX Denoise, and RTX Deblur for instant video post-refinement.
Model & LoRA Switches: Instant toggles for H3 Hybrid Model (FLV/REF), Turbo LoRAs, and NSFW LoRA.
Seamless Scene Extension (Continue Feature): Extend your video duration continuously with a single click. Currently, the extended video clip needs to be stitched manually in your video editor (Premiere, DaVinci, CapCut, etc.). Be sure to account for the overlapping frames for a perfect transition! (Full automated video concatenation will be integrated in a future update).
Audio & Preview Controls: Toggle real-time generation previews or mute embedded audio.
Performance Benchmark: Full 13.7s video generated in ~869 seconds (Latent Upscale Pass 1MP) on a single RTX 5060 Ti (16 GB).
Required Models & Custom Nodes
Required Models, Custom Nodes & Checkpoints
VAE:
Video VAE: minimax_h3_video_vae_fp16.safetensors
Audio VAE: minimax_h3_audio_vae_fp32.safetensors
Diffusion Models:
Hybrid Alternatives: H3 Eros Max | DaSiWa MiniMax H3
Text Encoders:
LoRAs:
Turbo 4-Step FL: minimax_h3_fl2v_turbo_4step_v1.2_768p_comfyui_bf16.safetensors
Turbo 4-Step REF: minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors
Ref Experimental: minimax_h3_ref_lora_rank_256_bf16
Motion & Reasoning: H3 - LTX 2.3 VBVR | Better Motion LTX/Minimax H3
Style: [MMH3] Mystic XXX (NSFW)
Latent Upscaler Model:
Custom Nodes
Make sure these node suites are installed via ComfyUI Manager:
ComfyUI-KJNodes
ComfyUI-VideoHelperSuite
ComfyUI-LogicUtils
Description
UPDATE v1.1 Changes
Visuals
- Greatly simplified and cleaner layout for better readability
Fixes
- Fixed first pass metadata/info handling
- Fixed Single Pass mode (already patched in v1.0b)
- Fixed Supersample / Upscale toggle logic
- Resolved various minor bugs
FAQ
Comments (6)
This is wonderful.
Just a little thing please :
How to add voice reference if the text is written
Tried joining audio file and prompting with " @audio1 is subject's voice 1 ..."
But impossible to have voice reference working.
Thanks
I just define it in
subject_defiitions:
@audio1 is the voice-timbre of <Subject 1>,
Then in retention_analysis I include:
@audio1: reference - timbre only; original words on the file are not spoken.
Then in shot_sequence:
<Subject 1> (S1) says, <d>[English] Hey how are ya!</d>
Not sure if this is the correct format but it works for me. (Total audio cannot exceed 15s)
Hope that helps
Sorry for late, and @123sirako123521 thanks for sharing this tip!
Best practice I found :
@image1 is Subject1, Helena.
@audio1 is Helena's voice; take its timbre, accent and pitch only, never its words.
Helena (S1) says, [English]: "Your new dialogue line here."
@123sirako123521 @SpaceGremlin Thanks for your tips !
Number 1: I love this workflow. The modular node being able to save and load setups is perfectly embedded in this and I love it.
But for the life of me I cannot get this Glide preview to actually preview anything. No matter what I try it jsut doesnt work. Is there anything in the requirements that are missing that are specifically for that to work? I can use the Model Preview Override and Sampler previews just fine in other workflows but they arent working here. Any help?
Hi, sorry, my bad! In the first pass (Get inputs), look for Get_Optimized_Model -> in the list, change "Optimized_Model" to "Model_Final". Do the same thing in SecondPass Latent Upscale -> Get Inputs -> Get_Optimized_Model. I'll release a fix for this soon.