Composition-locked storyboards to 4K with LTX-2.3 and RTX Video Super Resolution
NOTE
Install and dependency setup for each blueprint live in the GitHub repositories linked below. This playbook covers the creative workflow after those blueprints are installed. Concrete package pins, container tags, or a playbook-local asset clone are not included here.
Follow the setup instructions in each repository before you continue:
If you are new to ComfyUI, review How to Get Started With Visual Generative AI on NVIDIA RTX PCs.
Generate assets with the 3D Object Generation blueprint, either standalone or in Blender 4.2.
Standalone launch (PowerShell or Command Prompt), after you have followed the blueprint install path:
cd C:\3d-object-generation
conda activate 3dwithtrellis311
python app.py
When the app is running, open http://127.0.0.1:7860 and generate assets. Type a description of the scene you want to build (for example, “spaceship bridge”).
Run generation several times to build a collection of assets. You can also model directly in Blender or import props from elsewhere. Keep all assets in the same folder.
Open the sample Blender file that ships with the blueprint, remove the sample props and set decoration, then use the Asset Importer add-on to pull your content into Blender. You may need a scale factor (10x is a common starting point).
Camera angle, scene depth, and subject position in this layout carry through into the generated video.
Open Blender 4.5 LTS and open the 3D scene you built or edited. With the ComfyUI Blender AI node (ComfyUI x Blender) add-on installed from the guided blueprint, you should see it on the right side of the viewport.
Before you press Launch/Connect to ComfyUI, confirm the ComfyUI nodes are populated for both the first-frame and last-frame graphs:
unet Nameclip_name1, clip_name2sampler_name, schedulerPress the Launch/Connect to ComfyUI button on the add-on, wait 30–60 seconds for ComfyUI to load, pick your composition for the first frame, and press Run. The image saves according to the SaveImage node in the graph.
The graph builds a depth map from your Blender scene and combines it with your text prompt to generate a photorealistic image that matches your layout and perspective. Image generation uses FLUX.1 Depth, accelerated by NVFP4 on supported hardware.
Refine your prompt until the composition looks right. That image is your first frame.
In Blender, if the scene is animated, scrub to the end pose you want. If the scene is static, place a second camera and move objects to the end composition.
Then:
Press Run to create the last frame.
In ComfyUI, search the template browser for LTX and open the FirstFrame/LastFrame template. Load your first and last frame images into the corresponding input nodes. Write a video prompt that describes the motion between the frames as a short paragraph (natural language, not a tag list).
Example structure for a controlled camera move:
“Cinematic 1960s Supermarionation style. Two marionette pilots operate a retro cockpit… The camera performs a steady forward dolly-in… Outside, a rigid, static miniature space station… High-contrast studio lighting, visible model textures, and vintage 35mm film grain.”
Adjust CFG to change prompt adherence. Raising CFG (for example from 1 toward 4) aligns the generation more tightly to the prompt and can reduce creative variation.
A short negative prompt can help, though it is optional. Prompting tips: Prompting guide for LTX-2.
Iterate at 1280×704; when you like the motion, try 1920×1088. LTX needs pixel dimensions divisible by 32 (hence values such as 704 and 1088).
On supported hardware platforms, connect the RTX Video Super Resolution node to scale the output:
RTX in ComfyUI Manager and install the RTX Video Super Resolution nodeRTX in the Node Library and drag RTX Video Super Resolution into the graphFor 4K, choose 3× for a ~1280×720 source or 2× for 1920×1088. Stay on ULTRA quality unless you need faster performance.
When you are finished:
Cleanup is optional.