Prompt Studio MiniMax H3 Turbo Profile

Distilled "Turbo" LoRAs are the standard trick for making heavy video models usable on a real GPU - LightX2V is the best-known family of them, and MiniMax H3 has its own set of rank-reduced turbo LoRAs. The problem the pack author noticed is that these aren't plug-and-play: each one is coupled to a specific step count and a specific MiniMaxH3SigmaShift value, and the right file depends on which H3 mode you're running and what canvas you're generating at. Getting that wrong silently means 20+ steps on a model the distilled LoRA was never meant for, or mushy output. Prompt Studio MiniMax H3 Turbo Profile (node PSV_MiniMaxH3TurboProfile) is the node that picks the file and returns the matching settings, so you stop doing the lookup by hand.

How it works

It reads mode, width, and height - which is why it's designed to sit right after the pack's Director node, whose mode/width/height outputs are marked forceInput for exactly this. Then it applies one rule table:

• REF2VA always gets its task-specific 4-step LoRA (4 steps, video shift 12.0, audio shift 3.0).
• The exact 1344×768 training canvas gets the flagship FL2VA 768p 4-step profile (4 steps, shift 6.0) - under auto_quality.
• Any other base-mode canvas under auto_quality gets the mixed-resolution 8-step profile (8 steps, shift 12.0).
• fast_4step drops to the mixed 4-step profile away from 1344×768.

The model is patched through ComfyUI's LoraLoaderModelOnly, so no text conditioning round-trip is needed. Strength defaults to 0.75. There's an enabled toggle - flip it off and the node passes your model through untouched and reports the normal full-step settings (20 steps, shift 11.0/4.0), which is handy when you want to A/B the distilled path against the full model without rewiring.

Wiring it up

This is the one genuinely fiddly part, and the README is specific about it:

model output  -> your content LoraLoaderModelOnly chain -> MiniMaxH3SigmaShift
steps          -> BasicScheduler.steps
shift_video    -> MiniMaxH3SigmaShift.shift_video
shift_audio    -> MiniMaxH3SigmaShift.shift_audio

Put your own content LoRAs after the Turbo Profile's model output, not before - the profile owns inference-mode routing and the sampling policy, and content LoRAs should stay independent of both. The README's other hard rule: don't enable EasyCache or SpectrumApplyMiniMaxH3 in the initial Turbo workflow - their behavior on a 4- or 8-step distilled schedule simply hasn't been validated.

The LoRA files it wants

The four inputs fl2va_mixed_8step_lora, fl2va_mixed_4step_lora, fl2va_768p_4step_lora, and ref2va_4step_lora are file pickers that default to the exact Kijai LightX2V names (they live under ComfyUI/models/loras, typically in a MiniMax3 subfolder). Install them before queueing - the pack's Video Studio setup flow can download all four for you. The node resolves the configured file by its exact path, then falls back to an unambiguous basename match, so if you've moved a file into a different subfolder it usually still finds it.

Common issues

The most common error is a FileNotFoundError naming the missing LoRA in ComfyUI/models/loras - that's the four turbo files not being installed, not a broken node. If you have the same file in two subfolders, you'll get an "ambiguous" error telling you to pick the exact path in the dropdown. And the settings are coupled on purpose: don't hand-edit the steps/shift_video/shift_audio outputs downstream, because the profile returned them to match the LoRA it just applied. Start with auto_quality - it picks the 8-step mixed profile off the 768p canvas, which is the better quality/VRAM compromise - and reach for fast_4step only when you're genuinely GPU-bound. Distillation always trades a little quality for the speed; these profiles just make sure you're getting the speed you paid for.