From Still Image to Directed Scene — Your First Film with H3 CineDirector Global

From Still Image to Directed Scene — Your First Film with H3 CineDirector Global
A good still image already suggests a story. A lighthouse waiting in the mist.
A quiet room with sunlight moving across the floor. A character about to turn
toward the camera. The interesting question is: what happens next?
That is where ArtofTomworks H3 CineDirector Global comes in. It gives you a place
to describe the action, choose a camera movement and work with compatible native
H3 models inside ComfyUI. This guide takes you through a first scene, then shows
how to explore variations without rebuilding the workflow.
BEFORE YOU BEGIN
Download the complete CineDirector Global 1.1 package from its resource page.
The workflow JSON needs the included ArtofTomworks helper node; downloading
only the JSON is not enough.
You need a recent native-H3 ComfyUI installation, compatible H3 model weights,
the required text encoder and video/audio VAEs, rgthree and compatible KJNodes.
The animated preview also uses an H3 tiny preview decoder. Crystools is optional
and supplies the hardware readings in the upper monitor.
On Windows, run INSTALL_WINDOWS.cmd and point it to the ComfyUI folder containing
main.py. After any current job finishes, restart the ComfyUI ENGINE, refresh its
page and load ArtofTomworks_H3_CineDirector_GLOBAL.json. Restarting only the browser
does not reload newly installed Python nodes. CHECK_SETUP.cmd helps check the nodes.
Exact component filenames, download links and manual installation steps are in
DEPENDENCIES_AND_MODELS.txt and README_EN.txt. Model weights are separate downloads.
1. CHOOSE YOUR MODEL AND STARTING IMAGE
The included profiles target DaSiWa Hybrid V4 Turbo and Eros Max beta5. Choose your
installed model in the model panel. Its profile brings back its sampling settings,
components and saved LoRA stack while your scene, image, seed, duration and camera
stay in place.
The bundled settings belong to those exact variants, not every similarly named H3
model. For another compatible native H3 model, import a proven workflow or save an
explicit profile using your tested settings.
Replace the geometric technical sample with your own starting image. For this
example, use a coastal image with a lighthouse and visible water. A clear subject
and readable surroundings make the intended action easier to describe.
The standard resolution presets follow the image's portrait or landscape ratio,
with dimensions aligned to the model's size grid. Custom dimensions are available
when you want to choose width and height yourself.
2. DESCRIBE ONE CLEAR ACTION
In Scene and camera, replace the ENTIRE example scene prompt. Click inside its
text field, press Ctrl+A, then paste your new description. Try this with a matching
coastal reference image:
A photorealistic lighthouse stands above a rocky coastline at blue hour. Ocean
swells travel toward the shore, break against the rocks and spread into white foam
before receding. The lighthouse and cliffs remain solid and stable. Its warm beam
sweeps through a thin layer of sea mist. Natural motion and consistent exposure.
Notice what the prompt does: it identifies the subject, describes visible action
and says what should remain stable. You do not have to specify every possible
camera instruction here; the camera selector can add that separately.
For a different image, write the action that belongs to THAT scene. The coastal
example is a starting point, not a universal quality prompt.
3. MAKE A SIMPLE FIRST TAKE
For a first comparison, try:
• Camera movement: Locked camera
• Scene motion: Natural action
• Duration: 5 seconds
• Resolution: Quick — 768 long edge for a lighter test, or Balanced — 1024 long edge
Keep the model's profile settings as your starting point. If you are comparing
two camera or motion choices, keep the seed fixed so you change fewer things at once.
H3 uses a frame grid: a requested five-second shot becomes 124 frames at 24 fps,
approximately 5.167 seconds. The report shows the actual size and duration.
4. GIVE THE SOUND ITS OWN INSTRUCTION
The audio field is separate from the scene description. Replace its example too.
For the lighthouse scene, try:
Natural ocean surf and a gentle coastal breeze. No voices, no speech, no music.
Words matter here. If you want a desk fan, ask for a mechanical hum and airflow.
“Voice” means a human voice, so “fan voice” can send the request in the wrong direction.
These are generated sound instructions, not a library of guaranteed recorded effects.
5. RUN, WATCH AND REVIEW
Click Run. With compatible KJNodes and the preview decoder installed, the preview
can show animated glimpses during sampling. Treat these as approximations; judge
the finished video after decoding and saving.
The upper monitor shows status, elapsed time, progress and queue information.
With Crystools, it also shows hardware readings. The remaining-time estimate is
for the current sampling stage; decoding and saving can continue afterward.
Your MP4 appears in the result node and in your ComfyUI engine's actual output
directory under:
ArtofTomworks/H3_CineDirector/YYYY-MM-DD/
That is a relative subfolder inside YOUR configured output location. It does not
require the author's drive or installation path.
6. TURN THE SAME SCENE INTO A DIFFERENT SHOT
Once the basic motion makes sense, keep the image and scene prompt and change
one camera choice. Dolly in can bring attention closer to the subject. Dolly out
reveals more of the surroundings. A gentle orbit changes the viewing angle.
For our coastal example, try Drone rise with a Slow pace. Then compare it with
Drone forward flight, Drone orbit or Drone descent. The four drone choices add
different aerial instructions; they are not a physically guaranteed camera rig.
Start with modest movement, especially when the image gives little information
about what lies outside the frame.
If your scene prompt already specifies a complete camera move, From my prompt
prevents the camera selector from adding another one. Avoid asking for a static
camera in one place and a flying camera in another.
You can also try Loop-friendly scene motion for small cyclical actions. It is
guidance toward a returning movement, not automatic seamless-loop editing.
THREE SMALL HABITS THAT SAVE A LOT OF CONFUSION
Replace, do not accidentally append. Pasting at the cursor can leave the old
scene around your new text. During testing, an office scene and a leftover ocean
prompt produced a rather memorable indoor wave. Funny once; less useful when
you want a deliberate result! Replace the audio instruction at the same time.
Edit the scene in the Director. Actual generation prompt is the composed-text
display for inspection; it is not the scene editor. The Director feeds the H3
node through the connected prompt input.
Use the closing image deliberately. The default workflow uses one starting image.
A separate last-frame image can anchor an ending, but putting the identical image
at both ends can constrain movement. For a freer single-image test, leave the
last_frame input unconnected.
WHO SHOULD TRY IT?
If you already generate native H3 video in ComfyUI and want a more guided place
to direct the shot, this workflow is for you. Less experienced users can start
with a supported profile and the steps above, provided the H3 environment is ready.
It does not replace the initial model and node installation.
The interface supports English, Hungarian, German, French, Spanish, Italian,
Portuguese and Polish. Your scene text remains yours; generated camera instructions
stay in English. LTX, Ref2VA and GGUF pipelines need a different workflow branch.
For your first session, aim for one clear action and one readable camera move.
Once that works, change one thing and watch how the scene responds. That is where
the directing becomes fun.
MORE FROM ARTOFTOMWORKS
Explore more of my creative tools and workflows, available to download on GitHub:
https://github.com/artoftomworks-debug
If you enjoy these tools and want to support future development:
https://ko-fi.com/artoftomworks
Built with ComfyUI, KJNodes, rgthree and optional Crystools. Credits and dependency
links are included in the package. Share your results and tell me which camera
choice you tried first!