Multimodal to Video

These workflows feed one or more frames into the conditioning stream to steer the generated video. MiniMax H3 and LTX 2.5 offer different ways to combine a prompt with reference frames. See MiniMax H3 and LTX 2.5 for model details and dependencies.

Reference-to-Video

Reference-to-video conditions on reference frames to anchor subject identity and style. Encode the reference frame with the GGUF frame encoder and feed it into the conditioning stream alongside the text conditioning.

JSON fileMiniMax H3 Reference-to-Video workflow

Image-to-Video

A text prompt and one source frame produce the video. Load the source frame into the conditioning stream.

JSON fileLTX 2.5 Image-to-Video workflow

First and Last Frame to Video

A text prompt and two keyframe images produce the video. The model interpolates motion between the frames. Provide both keyframes to the conditioning stream.

JSON fileMiniMax H3 First and Last Frame to Video workflowJSON fileLTX 2.5 First and Last Frame to Video workflow

Models and formats

Model	Repository	Recommended formats
MiniMax H3	molbal/MiniMax-H3-GGUF	U16G, Q8_CR, FP8_Q4_0
LTX 2.5	molbal/LTX-2.5-GGUF	Q8_0, Q8_CR

Heaviest workflows in the registry

Dual-frame conditioning streams use the most memory. On 16GB cards, use U16G. See MiniMax H3 and Quant Formats.

Loading instructions

1. Download the GGUF files from Hugging Face.
2. Load the diffusion model, text encoder, and frame encoder with the GGUF loader nodes from comfyui-gguf-reboot. The loader nodes give the best performance.
3. Encode the reference or first-last frames into the conditioning stream.
4. Use a frame-aware latent and configure the video sampler.
5. Decode and assemble the frames into a video.

Showcase

Video examples are on molbal's Civitai profile.