Lock product identity
Keep the real device geometry, camera arrangement, buttons, logo, materials and colors stable before optimizing atmosphere or style.
Research · Workflows · Quality control
AI creation,
made repeatable.
A working notebook for practical AI-assisted video production. The focus is not just generating images, but building controllable workflows that preserve real product details and can be checked, reproduced and improved.
Local workstation baseline
These are the recorded environment facts. Installed models, custom nodes and successful end-to-end runs are tracked separately and are not inferred from the hardware specification.
RTX 5080NVIDIA GPU · 16 GB dedicated VRAM
~128 GBSystem memory, approximate
ComfyUIInstalled locally · model inventory is separate
Experiments and workflow records
Loading experiment records…
Working principles
Record dependencies
List model files, custom nodes, versions, inputs, outputs and parameters. Never assume a model or node is installed.
Define what “works” means
Evaluate visual consistency, temporal stability, resolution, artifacts and practical production usability.
Keep a rollback path
Preserve known-good workflows and document how to revert when a model update or dependency change breaks a result.
A practical image-to-video production path
Use this as a repeatable sequence for product advertising. It is a process template, not a claim that every model or custom node is already installed on the workstation.
Prepare a product reference set
Collect clean front, rear, side and detail views under neutral lighting. Keep one approved reference set and note any model-specific visual features that must not change.
Define one camera move
Write down shot duration, framing, camera direction, subject position and background action. Avoid asking the model to redesign the product and move the camera at the same time.
Validate a short sample
Start with a short, lower-resolution test. Change one variable at a time and record the seed, prompt, model, node versions and settings that produced the result.
Inspect identity frame by frame
Check silhouette, lens layout, buttons, ports, logo, material, color and scale. Reject clips where the product changes shape even if the overall image looks attractive.
Enhance only stable footage
After motion and product identity pass review, test enhancement or upscaling on a short segment before processing the full clip.
Keep a reproducible record
Save the workflow JSON, model file names, custom-node commit/version, input references, output sample, known issues and rollback instructions together.
Delivery acceptance checklist
- Product body shape and proportions remain stable throughout the clip.
- Camera modules, buttons, ports, logo and distinctive details match the reference.
- No obvious flicker, texture crawling, duplicated parts or frame-to-frame warping.
- Camera movement follows the shot plan and does not obscure key selling points.
- Output resolution, frame rate, duration and aspect ratio match the delivery brief.
- Model dependencies, settings and source assets are recorded for later reproduction.
Priority order: product identity → temporal stability → shot communication → image quality → stylistic polish.
Experiment record template
Copy this structure into the notes for each run. Mark unknown items as unknown instead of filling them with assumptions.
Experiment:
Date:
Goal / shot:
ComfyUI version:
Model files (exact names):
Custom nodes + versions:
Input references:
Prompt / negative prompt:
Seed and key parameters:
Output path:
Observed VRAM / runtime:
Product identity check:
Temporal artifacts:
Decision: PASS / RETEST / REJECT
Known issues:
Rollback / last known-good workflow: