Description

MiniMax H3, wired for you.

MiniMax H3 makes video and audio at once — but a working graph needs two different UNETs, a text encoder, two VAEs, a sigma-shift node, an advanced sampler, and a video muxer. Then you find out a single pass can't be long enough, so you start chaining clips by hand.

This is that whole graph, collapsed into one node.

Drop it on the canvas, point Settings at your models once, type a prompt, press Generate.

What you get

🎬 Three modes in one node Text only · First/Last Frame · Reference (up to 9 images + 3 videos + 3 audio, with in/out points)

🔗 Clip relay MiniMax H3 only accepts frame counts on a 17k+5 grid and one pass is VRAM-limited. Write more prompts, get more clips — each saved on its own, then stitched into one file

🎞️ Continuity that actually continues Last Frame Chain hands the previous clip's final frame over as the next clip's real first frame. Measured across the cut: 3.48 mean pixel delta, against 45.43 for unrelated frames

👀 Live preview Decoded frames stream into the node while it samples. You watch the shot form — not a spinner

🖼️ Gallery Every clip and stitched file in one place, each card keeping the prompt it was rendered from — one click to reuse or copy. Fullscreen player with Space / ← → / [ ] / Esc

✍️ Prompt workshopSplit a long brief into per-shot clips; the shared style preamble and sound tail are lifted into a common header/footer so every clip carries them. Optional Ollama LLM expansion, with the result shown for review before it lands

⚡ Acceleration SolAttn · Spectrum · Turbo LoRA (FL2VA) · H3 Cache · SageAttention — all selectable inline, all degrading gracefully when the pack isn't installed

🛡️ Guards, not crashes A mode whose UNET isn't set can't be entered, and the top bar says what's missing. VRAM is freed when the run ends

Continuity, in one table

Option What happens to clip 2 and onward Last Frame Chain Starts from the previous clip's final frame. Only FL2VA takes a first frame, so a continued clip renders on FL2VA whatever the run started as Reference (Reference mode only) Keeps rendering on Ref2VA with the same reference images — the look holds, the cut doesn't join None Nothing handed across; the run stays on its own model

Across all three, the common part of your prompt reaches every clip — that's what keeps a long piece looking like one piece.

Install

1 — The node

ComfyUI Manager → Install Custom Nodes → search TJ_NODE_STUDIO_ONE

or

cd ComfyUI/custom_nodes
git clone https://github.com/designloves2/ComfyUI-TJ_NODE_STUDIO_ONE.git

2 — Dependencies (one script, included in the package)

install_requirements.bat      ← Windows, double-click
./install_requirements.sh     ← Mac / Linux

Every MiniMax pack is optional — a missing one only switches its own feature off:

PackGives you

ComfyUI-KJNodes live preview

ComfyUI-SolAttn_triton SolAttn (the default accel)

ComfyUI-Spectrum-MiniMax-H3 Spectrum acce

lComfyUI-MiniMaxH3-Cache step-reuse cache

ComfyUI-MiniMax-H3-Turbo 4-step turbo (FL2VA only)

ComfyUI-VideoHelperSuite reference video inputs

Nvidia RTX Nodes RTX Video Super Resolution

pip install imageio-ffmpeg   # clip stitching needs ffmpeg

3 — Models — all from Comfy-Org/MiniMax-H3

→ folder file note

models/diffusion_models/minimax_h3_fl2va_

*Text only · First/Last · every chained clip

models/diffusion_models/minimax_h3_ref2va_

*Reference mode

models/text_encoders/qwen3vl_*_minimax_h3_

models/vae/minimax_h3_video_vae_* video VAE

models/vae/minimax_h3_audio_vae_* audio VAE

— must be the MiniMax onemodels/loras/minimax_h3*turbo*optional, FL2VA only

4 — Restart ComfyUI, open the node's ⚙ Settings, point each slot at your files,

press 💾 Save All. Done once, not per workflow.

Model licensing follows the upstream Comfy-Org / MiniMax repository — check there before commercial use.

Using this workflow

1. Load the .json

2. ⚙ Settings → pick your models → 💾 Save All

3. Pick a mode at the top of the node

4. Type a prompt. One prompt = one clip. Add prompts (or press ✂ Split into clips on a long brief) to make it longer

5. Set Continuity → Last Frame Chain if you want the clips to join

6. ▶ Generate

Node settings are saved inside the workflow as of v1.10.0 — this file opens the same way on someone else's machine.

🎁 And five more nodes come with it

Installing this gives you the whole ONE STUDIO family — the same one-node, no-wiring idea applied to images:

Node Model Modes

Z-Image ONE STUDIO

Z-Image TurboT2I · I2I · Inpaint · Outpaint · ControlNet · Face Redraw · RE-BG · Upscale

Flux.2 Klein ONE STUDIO

FLUX.2 Klein 9B / 4BT2I · I2I · Edit · Paint · Faceswap · Upscale

Qwen Image Edit 2511 ONE STUDIO

Qwen Image Edit 2511T2I · I2I · Edit · Inpaint · Angle · Faceswap · Upscale

Krea 2 ONE STUDIO

Krea 2T2I · I2I · IDENTITY 🧪 · ControlNet 🧪 · Upscale

SDXL ONE STUDIO 🧪SDXLT2I · I2I · Inpaint · Outpaint · Upscale

Same Settings panel, same gallery, same LoRA workflow, Korean / English UI. One install, six nodes.

📦 GitHub: https://github.com/designloves2/ComfyUI-TJ_NODE_STUDIO_ONE

Notes

• 🧪 The video node is experimental. Long pieces are assembled from chained clips, so they are not seamless.

• Generation time scales with clip count. The node measures your actual clip times and corrects its own ETA.

• Korean / English UI — switch in Settings.

Version name

v1.10.0

Version description

Continuity now genuinely continues: a chained clip renders on FL2VA with the previous clip's final frame as its real first frame (measured 3.48 across the cut vs 45.43 for unrelated frames). Splitting a brief keeps the shared style and sound text on every clip. Node settings are saved inside the workflow, so this file opens the same way on any machine. Gallery keeps each clip's prompt. VRAM is freed when a run ends.

Quick notes —

• One prompt = one clip. Add prompts (or use ✂ Split into clips) to make a longer piece.

• Want the clips to actually join? Continuity → Last Frame Chain.

• Reference mode needs the ref2va UNET; Text/First-Last need fl2va. Both go in Settings once.

• The audio VAE slot must hold the MiniMax audio VAE — another one fails at decode.

• Every accelerator pack is optional; missing ones just switch their own option off.

Bug reports and feature requests are very welcome on GitHub:

https://github.com/designloves2/ComfyUI-TJ_NODE_STUDIO_ONE/issues