Which Workflow Do You Need?
Pick the workflow that matches the input you already have.
| Your Input | Download This | |
|---|---|---|
| Only text | T2V workflow | Download ↓ |
| One image (first frame) | I2V workflow | Download ↓ |
| Two images (first + last frame) | FL2VA workflow | Download ↓ |
| Images / video / audio references | R2V workflow | Download ↓ |
| Faster on the same GPU | Turbo workflow | Download ↓ |
Workflow Library — Download JSON Files
Each workflow is a self-contained ComfyUI node graph. Download the JSON, drag it onto the ComfyUI canvas, and connect your model weights. Based on ComfyUI native MiniMax H3 templates.
Text-to-Video (T2V)
Generate a video from a text prompt alone. No image input needed. Best for creative ideation and exploration.
Requires: H3-Base FL2VA weights · ComfyUI (current build)
Image-to-Video (I2V)
Animate a starting image. The supplied image becomes frame 0 and the prompt describes what happens next.
Requires: H3-Base FL2VA weights · ComfyUI (current build)
First + Last Frame (FL2VA)
Upload an opening and an ending frame. H3 generates the visual transition between them. Best for controlled morph-style shots.
Requires: H3-Base FL2VA weights · ComfyUI (current build)
Reference-to-Video (R2V)
Use multiple image, video, and audio references to guide identity, motion, camera, or voice. Most flexible workflow.
Requires: H3-Base Ref2VA weights · ComfyUI (current build)
Turbo (Accelerated)
Run H3-Base in 4–8 steps instead of 25–40 using a Turbo LoRA. Large speed-up at a small quality cost.
Requires: H3-Base FL2VA weights · Turbo LoRA · ComfyUI (current build)
What's Inside a MiniMax H3 Workflow?
Every MiniMax H3 workflow file contains a ComfyUI node graph that chains together a standard set of processing stages. Here is the typical node layout for a T2V workflow.
| Node | Role |
|---|---|
| CheckpointLoader (H3-Base) | Loads the FL2VA or Ref2VA task-family weights |
| CLIP Text Encode (Prompt) | Encodes the positive text prompt |
| CLIP Text Encode (Negative) | Encodes the negative prompt |
| EmptyLatentVideo | Sets resolution, frame count, and aspect ratio |
| LoadImage | Loads first-frame and/or last-frame images (I2V, FL2VA) |
| KSampler | Denoising sampler (DPM++ 2M, UniPC, LCM for Turbo) |
| VAE Decode (Video) | Decodes latents to video frames |
| Audio VAE Decode | Generates synchronized stereo audio |
| SaveVideo (MP4) | Writes the final MP4 with embedded audio track |
I2V, FL2VA, and R2V workflows add image, video, and audio loading nodes plus their conditioning paths. The Turbo variant inserts a LoRA Loader and changes the sampler to LCM / DPM++ with a reduced step count.
Hardware Requirements
| Precision | VRAM Floor | Typical GPU | Notes |
|---|---|---|---|
| FP16 (standard) | ≥ 24 GB | RTX 4090 / 5090 / A100 | Best quality baseline. |
| FP8 | ≥ 16 GB | RTX 4080 / 4090 / 5090 | Common community setup. |
| INT8 / GGUF | ≥ 12 GB | RTX 4070 Ti / 4080 | Quantized for mid-range cards. |
| Apple Silicon | ≥ 32 GB unified | M2 / M3 / M4 Pro / Max | Community MPS builds; slower. |
VRAM usage is dominated by the H3-Base model weights and scales with resolution and duration. For detailed GPU-specific guidance, see the MiniMax H3 ComfyUI Guide.
Prefer to Skip the Install?
Same T2V, I2V, and reference workflows — running on hosted infrastructure. No GPU, no ComfyUI, no downloads. New users get 10 free credits daily.
Checking your account…
How to Load a Workflow in ComfyUI
Step 1 — Open the JSON File
Drag the downloaded .json file onto the ComfyUI canvas, or use Load from the toolbar menu. ComfyUI will render the full node graph.
Step 2 — Install Missing Nodes
If a popup lists missing nodes, open ComfyUI Manager, search for the missing node type, install it, and restart ComfyUI. Most H3 workflows use only built-in ComfyUI nodes.
Step 3 — Set Model Path and Generate
Confirm the CheckpointLoader points to your downloaded H3-Base weights (FL2VA or Ref2VA). Write your prompt, set resolution and duration, and click Queue Prompt.
For the complete install guide — including model file placement, offloading options, and SageAttention — see the MiniMax H3 ComfyUI Guide. For model weights, see MiniMax H3 Download.
Workflow Variants by Precision
Each workflow type can run at different model precision levels. Matching the right precision to your GPU is the fastest way to resolve out-of-memory errors.
| Precision | VRAM | Speed vs FP16 | Quality vs FP16 |
|---|---|---|---|
| FP16 (standard) | ≥ 24 GB | Baseline | Reference |
| FP8 | ≥ 16 GB | ~1.2× | ~99% |
| INT8 | ≥ 12 GB | ~1.4× | ~97% |
| GGUF | ≥ 10 GB | ~1.5× | ~95% |
Speed and quality estimates are approximate, based on community benchmarks. Actual results vary with GPU, driver, and ComfyUI version.
Workflow Troubleshooting
| Problem | Fix |
|---|---|
| "Missing Nodes" popup when loading JSON | Update ComfyUI to the latest build. Use ComfyUI Manager to install missing node packs, then reload the workflow. |
| Workflow loads but model is "not found" | Confirm weights are in the correct folder (diffusion_models, text_encoders, vae). Check filenames match exactly. |
| Out of memory during sampling | Lower resolution and video length first. Switch to FP8 or INT8 weights. Enable CPU offloading. |
| VAE Decode OOM after sampling finishes | Reduce output size. Increase system RAM. Try tiled VAE decode if the workflow supports it. |
| Generated video is all black / noise | Verify you loaded the correct checkpoint family (FL2VA for T2V/I2V, Ref2VA for R2V). Mismatched weights produce garbage. |
| No audio in the output | Check that the Audio VAE Decode node is connected and the audio VAE file is present in ComfyUI/models/vae. |
| Turbo LoRA has no speed improvement | Confirm step count is set to 4–8 (not 25+). Verify the LoRA loader strength is > 0. Match sampler to LoRA recommendation. |
For a comprehensive troubleshooting guide including advanced OOM strategies and SageAttention setup, see the MiniMax H3 ComfyUI Guide.
Workflow Version History
| Date | Version | Changes |
|---|---|---|
| 2026-08-18 | v1.2 | Added FL2VA first+last-frame and Turbo workflow variants. |
| 2026-08-09 | v1.1 | Updated to ComfyUI native MiniMax H3 templates. |
| 2026-07-15 | v1.0 | Initial release — T2V, I2V, R2V workflows. |
MiniMax H3 Workflow FAQ
What is a MiniMax H3 workflow?
A MiniMax H3 workflow is a ComfyUI node graph saved as a JSON file. It defines the full generation pipeline: model loading, prompt encoding, sampling, decoding, and output. You load it into ComfyUI, supply your prompt and any images, and run.
Where can I download MiniMax H3 workflow files?
The workflow library on this page provides free JSON downloads for T2V, I2V, FL2VA, R2V, and Turbo variants. The workflows are based on ComfyUI native MiniMax H3 templates.
Are these workflows official?
These workflows are based on ComfyUI native MiniMax H3 templates and community-tested configurations. MiniMax3.org is an independent platform, not the official MiniMax team. Official templates are available in the ComfyUI Template Library.
Do I need to download model weights separately?
Yes. The workflow JSON defines the node graph, not the model files. Download H3-Base weights (FL2VA or Ref2VA) from Hugging Face before loading the workflow.
Which ComfyUI version do these workflows require?
Use a current ComfyUI build with native MiniMax H3 support. Avoid pinning a specific version number — update ComfyUI, load the workflow, and let ComfyUI Manager resolve missing dependencies.
Can I edit these workflows?
Yes. ComfyUI workflows are fully editable node graphs. You can add LoRA loaders, swap samplers, change resolution, adjust steps, or rewire the pipeline freely.
What is the difference between the T2V and FL2VA workflows?
T2V takes only a text prompt. FL2VA adds first-frame and last-frame image conditioning — the same checkpoint handles both, but the node graph includes image loading and conditioning nodes.
How do I fix the "missing nodes" error?
Update ComfyUI to the latest build. If the error persists, open ComfyUI Manager, search for the missing node type, install the corresponding pack, and restart ComfyUI.
Can I use these workflows commercially?
The workflow JSON files themselves are freely reusable. Model weight usage is governed by the MiniMax license on Hugging Face. Check the license terms in the MiniMaxAI/MiniMax-H3 repository for commercial use.
What if my GPU cannot run any of these workflows?
Use the browser-hosted generator on this page — same H3 workflows running on hosted infrastructure with no install required. New users get 10 free credits daily.
Ready to Run MiniMax H3 Locally?
Grab the workflow that matches your input, download the model weights, and drag both into ComfyUI.
Go to Workflow DownloadsSources & License
ComfyUI MiniMax H3 documentation
Comfy-Org/MiniMax-H3 on Hugging Face (models)
Published 2026-08-18. Last reviewed 2026-08-18. Workflow JSON files are based on ComfyUI native MiniMax H3 templates and are freely reusable. Model weight usage is governed by the MiniMax license on Hugging Face. MiniMax3.org is an independent third-party platform and is not affiliated with, endorsed by, or operated by MiniMax.