DepriveD - MiniMax H3 All-in-One HIGH-SPEED Workflow

A simple, fast, high-quality MiniMax H3 workflow designed for consumer cards.

This is a dual-mode All-in-One:

• T2V / FFLF path

• Reference path (images + videos + audio)

A single switch at the top forces only one path to be active at a time so you can’t accidentally run both.

Core Features

• 7 + 3 two-pass latent upscale (using minimax_h3_latent_upscaler_3d_bf16.safetensors)

• Total of only 10 steps for high-resolution results that look like they were generated fully at the final resolution

• You get a quick first-pass preview you can examine before the second pass finishes — so you can see if the video is going in the direction you want without waiting for the full generation

• Optional FPS boost (leave at 24 to disable, or set to 30/48 etc.)

• Optional RTX Video Super Resolution on top (2.0× recommended, 1.5× also works well)

• Clean parameter panels — everything important is exposed, the rest is hidden

• Preset resolutions (all correctly divisible by 32 for the latent upscaler)

• Simple duration slider

• Optional LoRAs

• Full media loader for the Reference path (9 images / 3 videos / 3 audios) with resize, crop, trim controls inside the panel

How to Use (Quick Start)

1. Choose Resolution (or click the wheel to set your own — must stay divisible by 32)

2. Choose Duration

3. (Optional) Add LoRAs

4. Write your prompt

5. Hit Run

That’s it. Everything else has good defaults.

Recommended Starting Settings

• First Pass Steps: 7 (6 or 8 looks good too)

• Second Pass Steps: 3 (you can try 2 for a bit more speed)

• First Pass Denoise: 1.0

• Second Pass Denoise: 0.2–0.3

• Turbo LoRA Strength: 1.0 (lowering requires more steps)

• Video Shift: 12.0

• Audio Shift: 3.0

Increasing steps much higher with the Turbo LoRA will start burning the image. The starting values above are a strong baseline.

Important Notes

• Pictures do not save automatically. Right click any of the video outputs to save, OR set them to save instead of preview.

• The entire complex wiring is hidden behind the LoRA panels on each path.

• This is a new model and the community is still experimenting — feel free to try different settings.

• Between Sage Attention, the latent upscaler, and the Turbo LoRA, generation speed on a 5070-class card is competitive with (and in many cases faster than) typical LTX workflows while keeping strong quality.

Required Custom Nodes

• Pixaroma

• ComfyUI-VideoHelperSuite (VHS)

• ComfyUI-Fantastic-MiniMaxH3-PromptBuilder (Fantastic H3 Media Loader)

• VRGamedevgirl (Required for latent splitting and seperating.)

• Reroute nodes (usually included with base ComfyUI or common utility packs)

Required Models

• T2V-FFLF: minimax_h3_fl2va_pruned_int8_convrot

• Reference: minimax_h3_ref2va_pruned_int8_convrot

• CLIP: qwen3vl_32b_minimax_h3_nvfp4_awq (or your preferred quant)

• Video VAE: minimax_h3_video_vae_fp16

• Audio VAE: minimax_h3_audio_vae_fp32

• Turbo LoRA: minimax_h3_turbo_...

• Latent Upscaler (required): minimax_h3_latent_upscaler_3d_bf16.safetensors

Enjoy fast, high-quality MiniMax H3 generations!!!!

P.S.

PLEASE let me know anything I can do to IMPROVE or FIX the workflow, I'll update as fast as possible!