Bernini Updates Tile-Based Video Generation Workflow

Emotional_Example_12 · reddit · 2026-07-16

Bernini's FHD Tile Generation Solution

This post updates a video generation workflow for Wan / Bernini. The author replaced the legacy "Tile Settings" subgraph with a new Tile Split / Tile Select / Tile Merge 3-node setup, streamlining the process of splitting, selecting, and merging image/video tiles.

Key Improvements

Auto Grid Splitting: Automatically calculates tile dimensions based on grids (e.g., 2×2, 8×8, 16×16) and aligns them to the multiple of 16 required by the model.
Seamless Stitching: Tile Merge uses feathered overlap to combine processed tiles, avoiding visible seams.
Auto Upscale Detection: If a single tile is 2× the original, the merged image automatically scales to 2×.
Edge Padding: Resolutions like 1920×1080 (not a multiple of 16) are padded to 1088 and cropped back post-merge to prevent pixel loss.

Implementation Details

Pixel-based Tiling: Runs the entire pipeline independently on each tile, reducing global RoPE alignment issues.
Live Stitching: Reimplemented to fix tile drift by pasting adjacent tiles' generated outputs into overlap areas, tricking the model into treating them as finished.
Installer refactored to include Bernini-R INT8 ConvRot, LightX2V 4-step LoRAs, CUDA auto-detection, and idempotent downloads.

Bernini Features

Additional paper/experimental capabilities include:

Bernini Prompt Enhancer
First-Frame CoT
Bernini Multi-Guidance
guidancemode support in Infinity
Prompt Guide expanded to 22 Bernini-Bench tasks with 35 presets in total.

Finally, the author compares sequential vs. contextwindow video processing: the former saves VRAM, while the latter offers better temporal consistency and supports maskmode: bbox.

Original post →

Bernini Updates Tile-Based Video Generation Workflow

More from Multimodal