USE WHEN
Use this manual when
- Teams that need control over weights, dependencies, inputs and outputs
- Teams with Linux, NVIDIA CUDA and basic operations capability
- Projects requiring batches, private-network assets or custom nodes
PRODUCTION MANUAL · Local workflow
HunyuanVideo 1.5, Wan2.2 and ComfyUI: an open local production stack
Make a local video stack reproducible, portable and reversible across hardware, weights, nodes, workflow JSON and output metadata.
DO NOT USE AS
Do not use it as
- One-off projects without an environment owner
- Treating community plugins as official model support
01
VERIFIED FACTS
What the evidence currently establishes
- 01
The HunyuanVideo 1.5 repository publishes code and weights and lists Linux, Python 3.10+, a CUDA-compatible NVIDIA GPU and a 14 GB minimum VRAM figure with model offloading enabled.
- 02
The Wan2.2 repository publishes T2V, I2V, TI2V, S2V and Animate routes. Their support, dependencies and resolutions differ and must be verified per model.
- 03
ComfyUI documents human-readable JSON workflows and built-in templates based on core nodes that can prompt for required models; third-party nodes still need separate source and version locks.
02
REQUIRED INPUTS
Prepare before starting
- GPU, VRAM, driver, CUDA, PyTorch and OS inventory
- Model commit, weight filename, source URL and checksum
- ComfyUI version, core-node version and third-party node allowlist
- Fixed assets, prompt, seed, dimensions, frame count and sampling settings
- Output, metadata retention, logging and backup policy
03
WORKFLOW
Execution workflow
- 01
Build the minimum environment from official repositories before community bundles.
- 02
Verify weight hashes and record license, download date and commit.
- 03
Run a minimum sample with official or core-node templates and save the environment inventory.
- 04
Add custom nodes one at a time and rerun the same regression sample.
- 05
Bind workflow JSON, input hashes, seed and output metadata to the shot ID.
- 06
Clone the environment for upgrades; never overwrite production before validation.
04
DELIVERABLES & QC
Deliverables and acceptance
Required deliverables
- Locked environment and hardware inventory
- Weight provenance and checksum ledger
- Human-readable workflow JSON
- Minimum regression sample and expected output
- Upgrade and rollback record
QC checklist
- Clean environment rebuilds from documentation
- Output resolves to exact workflow and inputs
- Third-party nodes have source and commit
- Same regression sample runs before and after upgrades
- Production can roll back after failure
EVIDENCE BOUNDARY
What this manual does not prove
Minimum VRAM and speed statements apply only to the documented configuration, not every resolution, frame count or plugin mix. Community plugins, low-VRAM wrappers and bundles require separate provenance labels.