Difforum · Camera → Prompt

Prompt-driven video models don't take camera schedules. They take sentences. If you've drawn a careful crane-up-and-tilt-down in the Director, this node is what turns it into something MiniMax H3 or LTX will actually read - and it does it by using the pack's own move vocabulary, not by guessing.

Four output formats, one node

• sentence - one line of camera direction, plus the Director's look sentence if include_look is on. Append it to any prompt for any model. This is the default for a reason.
• timed lines - [0.0s-2.5s] camera ... per block. Good when the model handles time-stamped direction (and useful even if it doesn't, because you can read it and see whether your move is too busy).
• prompt suffix - the same thing framed as Camera: ..., for graphs that keep the camera clause separate from the subject description.
• H3 structured - the interesting one.

The H3 structured mode

MiniMax H3 has a native prompt format, and this mode writes your whole timeline into it: scenes become what's on screen, camera blocks become H3's own vocabulary (dolly_in → "pushes in", pan_right → "trucks right", crane_up → "pedestals up while tilting down to keep the subject framed"), keys become events at those moments, and the look opens [Shot 1] as the style.

h3_mode picks which task you're writing for: I2VA (first frame), FL2VA (first and last), or reference mode - which produces the six full-reference sections H3's ref2va path expects, with <Picture 1> as the first frame.

Two optional fields only exist in this mode and are worth filling in: soundscape (H3's overall_soundscape - ambience and physical sounds only; wind, footsteps, rain) and music (non_diegetic_music - instruments, tempo, dynamics, no mood words). H3 generates audio jointly with the picture rather than bolting it on, so these two fields are doing real work; leave them empty and you get soft natural ambience and no score.

cuts is the last dial: off means one continuous take where scene changes read as transformations, on means every scene starts a new [Shot N] at its cut time. For a music video, on. For a dream, off.

Inputs and output

direction gives it the timeline; camera overrides the camera if you want to describe a track that isn't in the Director. prefix is your shot description - what's on screen. In sentence and suffix modes it goes before the camera text; in H3 structured mode it becomes the opening composition, and in reference mode it's what <Picture 1> shows.

The output is a single text string, which goes into your CLIP text encode or straight into the video model's prompt field.

Install

Manager → search "Difforum", or:

cd ComfyUI/custom_nodes && git clone https://github.com/chillithebillis/Difforum.git difforum

Restart ComfyUI. Pure string assembly - no models, no dependencies. The H3 modes pair naturally with Template 10, which anchors frames that come from your camera move rather than ones you placed by hand.

Where people get burned

H3 structured raises an error if you don't connect a Director, because it needs the whole timeline - scenes and keys included - not just a camera. If all you have is a camera track, use sentence and prepend your own description.

And a note on the model side: H3's open weights ship under a community licence that excludes the US, EU, UK and South Korea from its applicable territory, and it generates 4–15 second clips natively. So a 30-second Director timeline becomes several H3 generations stitched end-to-end, not one prompt - which is what H3 Shot and the segment logic in Template 08 are for. Write the prompt, then split the shot.