Lazy MiniMax H3 — All-in-One (Auto-Switching)

Pick the workflow type. Everything else follows.

This is a Lazy MiniMax H3 graph: one selector, one conditioner path, one sampler stack. Set T2V / I2V / FL2V / R2V and the all-in-one nodes route models, frames, refs, and audio for you — no muted branches, no duplicate samplers, no “which UNET did I leave on?”

The Lazy idea

Lazy here means the workflow does the bookkeeping.

• You choose the workflow type once.

• Lazy MiniMax All-in-One builds the right conditioning for that mode.

• Lazy Model Switcher picks fl2va (text / image / first–last) or ref2va (reference).

• Lazy Image Loaders only emit what that mode needs; unused roles hard-gate off so optional sockets stay empty.

• LazyPrompt can expand a short idea into a MiniMax-ready prompt — then get out of the way.

Everything is modular. Nodes are grouped for clarity, not locked layout. Drag sections wherever you like; the wiring stays intentional.

Modes at a glance

ModeWhat you loadDiffusion

T2V

Text only

fl2va

I2V

First frame

fl2va

FL2V

First + last frame

fl2va

R2V

Reference images (+ optional audio)

ref2va

Change the selector → loaders, MiniMax, and UNET routing update together.

Prompt enhancement (optional, Lazy-friendly)

You can run raw text straight through (Prompt Engineer bypass) or enhance first.

Recommended: LM Studio (local LLM)

• Workflow can load the model, enhance, then unload when finished.

• Typical enhance time: about 5–10 seconds.

• Best quality/speed balance for day-to-day Lazy runs.

• Vision-capable LM Studio models work well with I2V first-frame grounding.

Alternative: TextGenerate (CLIP)

• Uses a wired Comfy CLIP / Qwen LLM — no second app.

• Works, but slow — often on the order of ~1 minute vs a few seconds with LM Studio.

• Fine if you refuse a separate LLM server; otherwise prefer LM Studio.

MiniMax-oriented skills (I2V / FL2V / R2V) are included so enhanced prompts speak H3’s language (<Picture N>, first/last alignment, reference jobs, soundscape fields).

What’s in the box

• Lazy Global Selector — single mode control

• Lazy Image Loaders — role + hard-gate by mode

• Lazy Subject & Scene Automation — optional subjects/scenarios/LoRAs

• LazyPrompt — Prompt Engineer — LM Studio / HF / TextGenerate CLIP

• Lazy MiniMax All-in-One — T2V ↔ R2V conditioner auto-switch

• Lazy Model Switcher — fl2va ↔ ref2va

• Native MiniMax H3 decode → video + stereo audio → save

Built for people who want one graph, not four copies of the same sampler.

Requirements (short)

• ComfyUI 0.30.0+ (native MiniMax H3)

• Vsaan212-workflow-utilities (Lazy nodes)

• KJNodes (Set/Get, helpers), plus usual Comfy core loaders/sampler/save

• Models from Comfy-Org/MiniMax-H3: fl2va + ref2va UNETs, Qwen3-VL encoder, video VAE, audio VAE

Quick start

1. Install custom nodes + download the H3 files into diffusion_models / text_encoders / vae.

2. Open the workflow; set Lazy Global Selector to your mode.

3. Drop images on the matching loaders only.

4. Write a short idea — enhance with LM Studio (recommended) or bypass for raw text.

5. Set duration on MiniMax All-in-One → Queue.