🍳 H3 Kitchen | FL2VA Video Extender v1.0

Two Clips • Two Prompts • One Continuous Video

🎬 Why I created this workflow

While experimenting with MiniMax H3 in ComfyUI, I wanted a simple way to generate videos longer than the usual single-clip duration.

Most of the longer-video workflows I came across were based on REF2VA, while I specifically wanted to use FL2VA (First/Last Frame to Video with Audio).

The goal was straightforward: generate the first clip, automatically extract its last frame, use that frame to start the second clip, and combine both clips into one longer video with audio.

No complicated interface. Just the essential controls, organized into a fun kitchen-themed workflow.

✨ Main Features

• 🎞️ Two consecutive FL2VA generations in one workflow.

• 🔗 Automatic last-frame extraction and handoff from clip 1 to clip 2.

• 📝 Two independent text prompts, allowing the action or story to change between clips.

• 🎬 Automatic video-frame concatenation.

• 🔊 Automatic audio concatenation.

• ⚙️ Shared controls for duration, FPS, sampling, seed, and resolution.

• 🧂 Optional LoRA support.

• 👨‍🍳 Organized, color-coded kitchen-themed sections.

• 📦 MP4 output with audio.

The included default settings aim for approximately 24 seconds of combined video, using two 12-second generation settings.

📖 How to Use

1. Kitchen Appliances

Load your MiniMax H3 FL2VA diffusion model, compatible Qwen3-VL text encoder, video VAE, and audio VAE.

2. Spice Rack

Select your LoRAs. A compatible Turbo LoRA is used in the example setup, with additional optional LoRAs available.

3. Recipe Measurements

Upload your starting image and choose the generation duration, FPS, sampling steps, sampler, scheduler, and other parameters.

Remember: the duration is applied to each clip separately.

4. Secret Recipe: The Important Part!

There are two independent prompt boxes.

Prompt 1: First Clip

Describe the starting scene and the action you want during the first clip.

Example:

• [0-4 sec] The chef prepares the ingredients.

• [4-8 sec] The chef begins cooking.

• [8-12 sec] The chef moves toward the serving counter.

Prompt 2: Second Clip

This prompt describes what happens after the first clip ends.

The last frame of clip 1 automatically becomes the starting frame of clip 2.

IMPORTANT: Reset the timestamps in Prompt 2!

Do not continue the timestamps from [12-16 sec].

Instead, write:

• [0-4 sec] The chef continues from the previous action.

• [4-8 sec] The chef plates the food.

• [8-12 sec] The chef presents the finished dish.

Although the story continues, the second prompt represents a new generation starting at zero seconds.

For better continuity, repeat essential details about the character, appearance, environment, lighting, and camera position.

5. GPU Cooking Station

Queue the workflow.

The first clip is generated, its last frame is extracted, and the second clip is generated using that frame.

Both clips are then combined automatically.

6. Takeaway Counter

The final video and audio are assembled and saved as an MP4 file.

⚙️ Example Settings

ParameterValueModelMiniMax H3 FL2VA INT8 ConvRotDuration12 seconds per clipFPS24SamplerEulerSchedulerSimpleSampling Steps8Video Shift8Audio Shift5OutputH.264 MP4

These are example settings from the workflow, not universal optimal values.

⚠️ Notes

• This version supports two sequential clips. It is not an unlimited-loop extender.

• The last frame provides visual continuity, but perfect motion, character, and audio continuity is not guaranteed.

• Users should select their own starting image and locally installed model files.

• Model loading and VRAM usage depend on the selected quantizations and available hardware.

• Install any missing custom nodes through ComfyUI Manager.

• On lower-VRAM GPUs, model offloading may be required.

🙏 Special Thanks & Credits

A big thank you to Stable Yogi and DaSiWa (Darksidewalker).

Their workflows, experiments, and contributions to the MiniMax H3 community helped me understand the available tools and develop this simpler setup.

Stable Yogi: https://stableyogi.com/

DaSiWa / Darksidewalker: https://civitai.com/user/Darksidewalker

This is a simplified, independently assembled two-pass FL2VA workflow inspired by the community's work.

🚀 Final Words

I wanted something simple enough to understand, customize, and experiment with without navigating a huge workflow.

This is my first public version, so feedback, suggestions, and improvements are welcome.

Happy generating, and enjoy your time in the H3 Kitchen! 🍳🔥

PLEASE LET ME KNOW IF THERE ARE ANY WRONG WITH THE WORK FLOW TO FIX IT