Bernini Run Single GPU

This is the centerpiece of the pack, and the reason the other Bernini nodes exist. Bernini is a 14B renderer with its own Python 3.11 / CUDA 12.4 requirements, and bolting that onto ComfyUI's environment is a recipe for a broken install. Bernini Run Single GPU sidesteps the whole mess: it launches the upstream infer_single_gpu.py in a subprocess, streams the logs into your console, and lets ComfyUI keep running normally even if Bernini's dependencies aren't installed in ComfyUI's Python at all.

How it works

The node assembles a command - your configured Python, the repo's infer_single_gpu.py, --config, --case, and every generation flag - writes the case to a temporary JSON file, runs it with a timeout, and reaps the process if it hangs. Relative output paths are resolved under ComfyUI's output directory, output folders are created automatically, and a small progress bar ticks in the corner. It's the Kijai/WanVideoWrapper philosophy applied to a CLI script.

Prefer the typed inputs

Three optional inputs override the wall of fallback widgets:

• bernini_config (from Bernini Model Config) - repo path, Python, config dir, checkpoints, env vars.
• bernini_case (from Bernini Case Builder) - prompt, task type, guidance mode, media paths, output.
• generation_args (from Bernini Generation Settings) - frames, size, steps, seed, omegas, eta.

Connect all three and you can ignore most of the widgets. The fallback widgets exist so you can test a one-off prompt without building the whole graph.

The controls worth knowing

• dry_run - builds the full command and case but doesn't execute it. This is the most useful toggle in the node: it's how you verify your paths and flags before committing to a 30-minute render. Do this every time you change something.
• timeout_seconds - default 1800. Hung processes get killed and reaped instead of piling up.
• use_prompt_enhancer - passes --use_pe, which needs BERNINI_PE_API_KEY in the env (set it in Bernini Model Config's env_json).
• negative_prompt / system_prompt - optional extra conditioning passed as CLI flags.
• extra_args - raw CLI flags for upstream options not yet exposed; shlex-split with Windows handling.
• show_preview - for image outputs inside ComfyUI's output dir, returns UI preview metadata so the result appears in the node.

Outputs

• output_path (STRING) - where the file landed. For .png/.jpg/.webp results, feed it to Bernini Load Output Image to get a real IMAGE tensor. For .mp4, keep the path for video helpers.
• case_json (STRING) - the exact case that was run.
• command (STRING) - the exact command line, quoted and shell-safe. Invaluable when a render fails: you can run it by hand in the Bernini environment and see the real error.
• log (STRING) - the streamed subprocess output.

Gotchas

• The repo must exist - if infer_single_gpu.py is missing you get a clear "clone ByteDance/Bernini first" error, not a traceback.
• Editing is slow by design: Bernini is Wan-2.2-derived, and the source video you're editing is part of the model input, so compute is roughly double a plain Wan generation. On consumer GPUs, budget accordingly (a single 3090 was reported around 55–60 s/it at 480×832/101 frames).
• The two-checkpoint rule follows you here too: high_noise_ckpt/low_noise_ckpt must be set together.

Installing this node

One of the 16 nodes in the lzits Nodes pack (Lior Zitser, MIT). Install via ComfyUI Manager (search "lzits Nodes") or:

cd ComfyUI/custom_nodes
git clone https://github.com/lzitser23/lzits-nodes.git

Restart ComfyUI. The pack itself needs no Bernini Python deps, but you still need the upstream pieces: clone bytedance/Bernini, download ByteDance/Bernini-R-Diffusers (a real download), and have a Python 3.11 CUDA 12.4 environment. The Bernini Setup Commands node prints the exact commands.