This ComfyUi fl2v workflow use Turbo 4steps with Latent Upscaling to generate 1.0 mp good quality result with short generating time.

If this workflow works great for you, please support me with ⚡⚡buzz⚡⚡, so I can do more testing and share more works, thank you!🦉❤️☕

Generating Pipeline:

1. Turbo generating with 0.2mp in the first 2 steps.

2. Upscale the H3 AVLatent to 1.0mp using trained neural network.

3. Re-sample the H3 AV latent at higher resolution in the last 2 steps.

This ensures decent generating speed + Low VRAM requirement + Naturel fine upscaling details.

Features:

• 🚀Turbo 4 steps generating (2+2)

• 🔍Latent Upscaling

• 🗒️Multiple LoRa support

• 🖥️Runtime Preview when generating

• 🗜️GGUF models

Tested on RTX3080(laptop) 16GB VRAM .

• Choose Q8 Base model and Q5 text encoder if you have 16GB VRAM.

• Choose Q4 Base model and Q4 text encoder if you have less than 16GB VRAM.

Model Links

- Base model -
minimaxH3TurboGGUF_fl2vTurbo4stepV10768.gguf(Q8)

• place it under: models\unet\

• Choose Q8 for 16GB+ VRAM, Q4 for less

Thanks molbal for his awesome model!

- Text Encoder -
MiniMax-H3-encoder-Q5_K_M.gguf(Q5)

Make sure also download: MiniMax-H3-encoder-mmproj-F16.gguf from that page and place it under the same folder.

• place them all under: models\text_encoders\

• Choose Q5 for 16GB+ VRAM, Q4 for less

Thanks joeygambino for his awesome model!

- VAE -

Video VAE

Audio VAE

• place them all under: models\vae\

- Preview VAE -

Tine VAE

• place it under: models\vae_approx\

Thanks Kijai for his awesome model!

- Upscaler -

minimax_h3_latent_upscaler_3d_fp16.safetensors

• place it under: models\latent_upscale_models\

Thanks LBH-123-AI or his awesome model!

ComfyUi Custom Nodes

- ComfyUI-H3-Multishot

Used for GGUF base model loader, GGUF text encoder loader, Multi LoRa loader,

Github install URL: https://github.com/jlucasmcrell/ComfyUI-H3-Multishot

- Comfyui_Minimax_h3_latent_Upscaler

Used for Latent upscale

Github install URL: https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler

Make sure update your ComfyUi to latest version!

Tips

- How to use Last Frame -

• Connect resized image (0.2mp) to last_frame of MiniMax H3 Image to Video

• Connect resized image (1mp) to last_frame of Upscaler Condition

- How to switch to [t2v] -

• Bypass the following nodes: First Frame, Resized Image (0.2mp), Resized Image (1mp)

- Balance between motion and surface details -

• Lower the Shift_video value of ModelSamplingMiniMaxH3 node will make the model focus more on physics, increase it will make it focus more on surface details.