lzits-nodes

A collection of custom nodes for Comfyui

Included Nodes

• String Splitter
• String Appender (Suffix)
• String Prepender (Prefix)
• Index Picker
• LoRA Selector Multi-Output
• Image Outpaint (Color Canvas)
• Bernini Model Config
• Bernini Generation Settings
• Bernini Case Builder
• Bernini Run Single GPU
• Bernini Conditioning (Kijai Branch)
• Bernini Load Output Image
• Bernini Setup Commands

Image Outpaint (Color Canvas)

This node lets you:

• Upload/load an image directly in the node UI.
• Expand the canvas on left/right/top/bottom without scaling the original image.
• Fill new outpaint regions with a chosen color (#RRGGBB, #RGB, r,g,b, or basic names like white).
• Output a mask where 1.0 is the new outpaint area and 0.0 is the original image.

Bernini nodes

Adds first-pass ComfyUI wrappers for ByteDance's Bernini-R renderer from https://github.com/bytedance/Bernini.

The layout now follows the Kijai/WanVideoWrapper style more closely: small typed helper nodes feed a runner node, heavy model code is not imported at ComfyUI startup, widgets have tooltips/descriptions, and the subprocess runner streams logs instead of waiting silently.

Recommended graph: Bernini Setup Commands -> install upstream assets manually, then wire Bernini Model Config + Bernini Case Builder + Bernini Generation Settings into Bernini Run Single GPU. For one-frame/image results, connect the runner's output_path to Bernini Load Output Image to get a normal ComfyUI IMAGE tensor.

Bernini Setup Commands

Outputs copy/paste setup commands to clone Bernini, install its requirements, and download ByteDance/Bernini-R-Diffusers. Bernini currently wants Python 3.11, CUDA 12.4, PyTorch 2.5.1+cu124, and a strong CUDA GPU; H100-class hardware is recommended by the upstream repo.

Bernini Model Config

Bundles repo path, Python executable, model config directory, optional high/low checkpoint paths, and optional env vars into a typed BERNINI_CONFIG object. This mirrors Kijai's loader/config nodes: set model/env paths once, connect them into the runner, and keep the runner less cluttered.

Bernini Generation Settings

Bundles frame count, size, steps, seed, fps, guidance mode, and Bernini omega/eta/flow-shift values into a typed BERNINI_GENARGS object.

Bernini Case Builder

Builds a Bernini case JSON object with task_type, guidance_mode, prompt, media paths, and output path. It outputs both plain strings and a typed BERNINI_CASE object for cleaner graphs.

Bernini Run Single GPU

Runs upstream infer_single_gpu.py in a subprocess so ComfyUI can still start even if Bernini dependencies are missing. Prefer connecting:

• bernini_config from Bernini Model Config.
• bernini_case from Bernini Case Builder.
• generation_args from Bernini Generation Settings.

Fallback widgets are still present for quick one-node tests. The runner supports dry_run to build the command/case without launching Bernini, plus extra_args for upstream CLI flags that are not exposed yet. Relative outputs are resolved under ComfyUI's output directory when available, output folders are created automatically, and timeout handling kills/reaps hung upstream processes.

Bernini Load Output Image

Loads a generated .png, .jpg, .jpeg, or .webp output path into a ComfyUI IMAGE tensor so normal Preview/Save Image nodes can consume Bernini t2i/i2i results. Leave video .mp4 outputs as path strings for video-specific helper nodes.

Bernini Conditioning (Kijai Branch)

Optional native-conditioning bridge based on Kijai's ComfyUI/tree/bernini branch. It accepts positive/negative conditioning, a VAE, optional source/reference video/images, and emits context_latents plus a matching Wan latent. This only changes model behavior when your ComfyUI build understands context_latents (Kijai's Bernini branch or a future upstream equivalent); on normal mainline ComfyUI it is present for compatibility but the core Wan model may ignore that extra conditioning.

For text-to-image, use task_type=t2i, num_frames=1, and a .png output path. For video, use one of the video task types, increase num_frames, and use a .mp4 output path.