EXAMPLES.md

Recipes

Ten short projects, each a template plus a shot script from example_scripts/. Load the template, put a Shot Script node on the Director's timeline_in (template 16 has one wired) and type the script name in file, for example examples/01_infinite_zoom.txt. Or open the .txt, read it and draw the same beats by hand. Every script is plain text: copy one, change the words, queue.

#	recipe	template	script	what it shows
1	Infinite zoom	02 / 05	01_infinite_zoom	the Deforum classic: scenes morph as the camera falls in
2	Music video on the beat	04	02_music_video_beat	camera blocks that react to the track (pulse, shake, bass speed)
3	Seamless loop for VJ / LED	06	03_seamless_loop_vj	camera and prompt close on themselves
4	A still with real depth	03	04_parallax_still	2.5D move on a photo or key visual, low energy keeps the picture
5	Cinematic move on H3	08	05_cinematic_push_in	one frame, a drawn camera, H3 does the motion
6	Storyboard to shot	16	06_storyboard_to_shot	your stills as keyframes, no look pass
7	AnimateDiff look on footage	15	07_animatediff_dance	AnimateLCM over a clip you shot; the script is a list of styles to try
8	Deforum morph over any clip	13	08_deforum_morph_restyle	the look changes over time, the motion stays the clip's own
9	One-minute installation shot	12	09_installation_long_shot	a long take with key moments that must land
10	Previz and camera hand-off	01, 09	10_previz_camera_export	block, show and export to After Effects / Blender with no model

Three stories

Templates 17-19 are finished short pieces to open, queue and then rewrite. They avoid the things that make AI video feel off: no faces, no spirals, one look held from the first frame to the last.

template	engine	why it holds together
17 · Story in three shots	Scene Stills → MiniMax H3	the image model paints three related pictures, H3 only moves them, scenes are cuts instead of morphs
18 · Living photograph	Feedback Sampler + depth	energy 0.2-0.34: geometry stays, the light changes; a 3D camera on a depth map
19 · Painted story	Feedback Sampler, turbo	one subject that grows, so each change reads as story; 12 fps, the look repeated in every scene

What to keep when you write your own:

• One look, written once and repeated. Scene Stills has a style box; in feedback templates the look words open every scene prompt.
• Energy is the realism dial. Under 0.3 the picture moves, around 0.4 it is re-painted, over 0.5 it morphs. Raise it only on the beat that changes.
• Slow, simple cameras: push in, pan, rise, at speed and amount 0.4-0.7. Rolls, spirals and vortexes are what reads as psychedelic.
• Change one thing per scene - the light, the weather, the size of the subject - and keep the subject words identical.
• Objects, places, hands and silhouettes survive small render sizes; faces do not.

Writing the script with an LLM

Shot Script's llm_instructions output is a prompt that asks for this format with the clip length and the valid camera moves. Feed it, plus one line about the film, to any text node, and wire the answer into script_in:

LLM node (prompt = llm_instructions + "A storm arrives at a lighthouse.")
   -> Shot Script.script_in -> Director.timeline_in

Set the Director's external to text only to keep a camera you drew and take only the words, or replace to let the script direct everything.

Fast first, quality last

A long render is almost always one of four things. Reference figures from a 12 GB card: about 2 s per diffused frame (SDXL turbo, 4 steps, 640 px) and about 14 s per MiniMax H3 step (640 px, 124 frames).

what costs	fast setting	what it saves
H3 at 20 steps	Turbo LoRA at 4 steps (on in the templates)	5x on the H3 pass
H3 Latent Upscale (refine at 2x)	off while exploring, on for the final	about half the H3 time
Restyle on every frame	off while exploring; cadence 2 when on	one image-model pass per frame
a 4x upscale model on large frames	Upscale model_use = auto skips it under 2x	most of the upscale time

Rough budgets with those settings:

shot	route	diffusions	order of time
10 s Deforum look, 24 fps	template 02 / 05, turbo model, cadence 2	120	~5 min
60 s Deforum look, 12 fps	template 12, turbo model, cadence 3	240	~10 min
5 s H3 shot	template 08 / 16, Turbo LoRA	4-8 H3 steps	~1-2 min + model loading
5 s H3 with the look	template 11 as shipped (look pass cadence 2)	62 + 4 H3 steps	~4 min + model loading

For long pieces:

• Feedback look: render at 12 fps and 640 px with cadence 2-3, then Upscale 2K; interpolate to 24 fps with any frame-interpolation node if the piece needs it.
• H3: the model is trained on 5-15 s clips. Split a long shot into segments with H3 Shot (segment), each starting on the last frame of the one before.
• Launch flags: --cache-none makes every Queue run the whole graph again, including text encoding and passes that did not change; without it, changing only the Restyle re-runs only the Restyle. --disable-smart-memory unloads the image model between the VAE and the sampler; Difforum suspends it while a frame is made so both stay loaded (set DIFFORUM_RESPECT_MEMORY_FLAGS=1 to opt out).
• Previz first: the Animatic costs seconds. Switch Render off (or press Previz only on the Director) until the timing is right.