a reference image for character identity and appearance
a source video for pose, action, timing, camera choreography, depth, composition, Canny/edge layout, and scene geometry
DiffusionGemma prompt synthesis through image_identity_video_control
Canny, Depth, and DWPose control branches for comparison

The workflow uses exactly these DiffusionGemma nodes:

DiffusionGemmaModelLoader
DiffusionGemmaContextHub
DiffusionGemmaTargetProfile
DiffusionGemmaCoTGenerator
DiffusionGemmaJSONSplitter

Bundled media lives under:

examples/assets/character_motion_transfer

Before running the workflow, copy that folder to:

ComfyUI/input/character_motion_transfer

Expected workflow inputs after copy:

character_motion_transfer/character_reference.png
character_motion_transfer/motion_control_video.mp4

Included demo output:

examples/assets/character_motion_transfer/diffusiongemma_ltx_character_demo.mp4

Wire downstream generation nodes from the JSON Splitter outputs only when ready_for_generation=true.

MiniMax H3 Director T2VA

The native text-to-video/audio example is:

09_minimax_h3_t2va_director.json

It converts a normal creative brief into Director's dedicated minimax_h3_prompt, wires the splitter's width and height outputs into native MiniMax-H3, and shares one editable duration between the Director profile and H3 frame calculation. Parallel previews show the generated H3 prompt and the splitter metadata/blocked reasons without interrupting the direct prompt connection. The included preset is 20 seconds, 481 frames at 24 fps, 0.4 MP at multiple-of-32 resolution, and 25 sampling steps.

The workflow uses the native MiniMaxH3ImageToVideo path plus RHMiniMaxH3SageAttentionPatch; install ComfyUI_RH_MinMaxH3 and place the four model files selected in the embedded H3 group in their normal ComfyUI model folders. DiffusionGemma is set to unload after prompt generation so H3 can reclaim VRAM before sampling.

MiniMax H3 Director Ref2VA

The ordered-reference example is:

10_minimax_h3_ref2va_director.json

It uses two images in explicit <Picture 1> / <Picture 2> order, analyzes that same ordered batch with DiffusionGemmaH3ReferenceContext, and sends the dedicated Ref2VA prompt directly to native MiniMaxH3ReferenceToVideo. The sample creative brief explicitly requests 2D hand-drawn animation and assigns protagonist identity to Picture 1 and environment/style to Picture 2, preventing vague “inspired” wording from drifting into generic CGI subjects.

One editable duration is shared by Director planning and native frame calculation; the included setting is 20 seconds, 481 frames at 24 fps, with 25 sampling steps. The Director uses its full 2048-token output budget because the official Ref2VA representation includes subject definitions, retention analysis, and a detailed timeline in addition to audio sections. Replace both example.png placeholders before running and keep their socket order aligned with the reference manifest. The Director loader unloads before H3 sampling, and both original images feed the native Ref2VA node separately.

LTX 2.3 Reframe with selectable audio

The provisional, unexported repository prototype is:

gemmaREFRAME_Experimental_audio_on-off.json

The current package entry point does not register the LTXReframe* nodes, and the retired layout frontend is not shipped. Keep these reframe artifacts separate from the stable Director examples when preparing a Registry release.

Its Use Custom Vocal Audio? switch defaults to off:

Off: the source video's audio is used for both LTX audio conditioning and the final video. Videos without an audio track use duration-matched silence.
On: upload a custom audio track, choose its start time in Custom Track Window, and use the MelBandRoFormer vocal stem. Long tracks are trimmed and short tracks are padded with silence to match the effective video range.

Required custom-node dependencies:

ComfyUI-MelBandRoFormer
comfyui-kjnodes
comfy-mtb

Required model:

ComfyUI/models/diffusion_models/MelBandRoformer_fp16.safetensors

No custom music is bundled. Source-audio mode can remain selected without choosing a custom file; custom-audio mode reports a clear error until an audio file is uploaded or selected.