Turbo + lower VRAM: reduce 20 steps to 4–8 — MiniMax H3 tutorial
Add the community Turbo nodes to a working H3 ComfyUI baseline, render usable clips at 6–8 steps, and follow a controlled fallback order when VRAM runs out.
Audience: NVIDIA users who can already render with the official H3 graph but need faster iteration or lower VRAM use.
Hardware: An NVIDIA GPU. Record native peak VRAM and time first. Low-VRAM mode is a fallback, not the default quality setting.
Prerequisites
- A stable native T2V or I2V workflow from the first guide
- A fixed seed, prompt, resolution, duration, and input for A/B testing
- ComfyUI-Manager or the ability to install custom_nodes manually
- minimax_h3_turbo_v4_step600_ema.safetensors from Larryvrh's Hugging Face repository
Steps
- Run the native 20-step graph once and record total time, peak VRAM, fine detail, fast motion, and audio quality. This render is the only valid baseline.
- Install MiniMax-H3 Turbo through ComfyUI-Manager, or clone it under ComfyUI/custom_nodes/. Fully restart ComfyUI and confirm there is no import error.
- Place minimax_h3_turbo_v4_step600_ema.safetensors in ComfyUI/models/loras/. Prefer v4-600 for most work; compare the older v1-850 only when 4-step heavy motion shows trailing or smear.
- Start from the official T2V or I2V graph. Insert MiniMax-H3 Turbo LoRA between Load Diffusion Model and SamplerCustomAdvanced, then select the downloaded checkpoint.
- Feed SamplerCustomAdvanced from MiniMax-H3 Turbo Sampler and set BasicScheduler to simple. Keep conditioning, both VAE decoders, and audio output unchanged.
- Start at 8 steps and strength=1.0, then test 6 steps. Use 4 steps for composition previews or strict budgets rather than assuming it is final quality.
- Render again with the exact same seed, prompt, media, resolution, and duration. Compare time, face and hand detail, fast-motion trails, audio quality, and sync. Turbo must be the only changed variable.
- If VRAM runs out, reduce preview resolution or duration first, disable nonessential previews or nodes second, and enable low_vram last. Low-VRAM mode lowers peak use but can soften quantized output.
- For smear or blur on a specific clip, test strength around 1.05–1.2. For oversharp grain, test about 0.8–0.95. Change only one parameter per run.
- Keep separate native and Turbo workflow files. Use Turbo for shot selection and bulk previews, then compare important finals against the native 20-step baseline.
Commands
cd ComfyUI/custom_nodes
git clone https://github.com/larryvrh/ComfyUI-MiniMax-H3-Turbo
git -C ComfyUI-MiniMax-H3-Turbo pull
ComfyUI/models/loras/minimax_h3_turbo_v4_step600_ema.safetensors
Caveats
- The useful range is 4–8 steps, with 6–8 usually much better than 4. Going above 8 adds no reliable benefit and may oversharpen.
- Keep strength at 1.0 by default and tune only for a specific smear or oversharp failure.
- low_vram trades sharpness for memory and should stay off when memory is sufficient.
- Audio and intense fast motion remain active areas of improvement. Fewer sampling steps are not the same as proportional end-to-end speedup.
- Do not stack Turbo, Spectrum, EasyCache, and other optimizers by default; their errors and audio losses can compound.
View original source · Larryvrh · 2026-08-23