Bernini-R Workflow · ComfyUI-BerniniR WrapperA node pack for the Bernini-R (1.3B / 14B) video generation model — VRAM optimization · multi-mode guidance · context window · differential-diffusion masking · segmented prompt control github：<a target="_blank" rel="ugc" href="https://github.com/xiaolibai-sys/ComfyUI-BerniniRWrapper">xiaolibai-sys/ComfyUI-BerniniRWrapper</a> <h1 id="overview">Overview</h1>This workflow is built on the ComfyUI-BerniniR Wrapper node pack. A single graph combines VRAM optimization, a multi-mode guidance family, temporal context windows, differential-diffusion masking, and segmented prompt control — letting long videos run on consumer GPUs / low-RAM machines while still allowing fine-grained per-segment content control.<h1 id="1-vram-and-loading-optimization">1 · VRAM &amp; Loading Optimization</h1>● Block Swap Per-layer DiT weight swapping between GPU and CPU with pin_memory and asynchronous prefetch, drastically cutting peak VRAM so the 14B model runs on limited memory.● Streaming Loader Optimized for low-RAM devices — weights are streamed from disk on demand instead of loading the whole model into memory at once.● Automatic Model Offload Lazy-loading strategy: weights are loaded only when used and released when idle, avoiding two copies of the same model weights in memory and keeping the memory peak low.<h1 id="2-guidance">2 · Guidance</h1>A seven-mode guidance family is included: CFG, APG, RAAG, S2, Z2, STG_A, STG_R — covering classifier-free, projected, ratio-aware, stochastic self-, zero-cost zigzag, and spatiotemporal skip guidance. A per-step guidance strength schedule (cosine / linear / piecewise) is also supported. See the separate guidance documentation for details.<h1 id="3-context-and-noise-control">3 · Context &amp; Noise Control</h1>● Context Window Paired with FreeNoise noise shuffling, long videos are split into overlapping temporal windows to keep motion and semantics coherent across segments and avoid long-sequence degradation.<h1 id="4-conditioning-and-masking">4 · Conditioning &amp; Masking</h1>● Differential Diffusion Mask A differential-diffusion mask mechanism supporting freeze / anneal on selected regions — useful for inpainting, reference-image preservation, or regional conditioning.● NAG (Normalized Attention Guidance) Injects a negative-prompt attention path to enhance detail.<h1 id="5-sampler">5 · Sampler</h1>● Dual Expert Sampler Switches between high-noise and low-noise expert models by split step, balancing early structure formation with late detail convergence.<h1 id="6-segmented-prompt-control">6 · Segmented Prompt Control</h1>● BerniniR SegmentSchedule Generates differing prompts between windows for per-segment semantic choreography.<h1 id="prompt-format">Prompt Format</h1>1-32(index1): prompt1 ; 33-64(index2): prompt2Each segment is tagged with a frame range + index, separated by semicolons; the index maps to a window in SegmentSchedule.<h1 id="recommendations">Recommendations</h1>● Auto windowing The workflow windows automatically — no manual Context Window configuration needed.● Recommended: continuous motion control Windows share local context, ideal for character action, camera moves, and other strongly continuous tasks.● Use Full load when RAM is ample If your machine has enough RAM, prefer the Full (eager) load mode over the streaming loader — weights stay resident in memory so startup and generation begin faster; the streaming loader only pays off on memory-constrained systems.Note: storyboard / multi-scene shots perform worse — because local context is shared, large semantic jumps between segments are weak. For those, use an explicit Context Window or separate batches.<h1 id="limitations">Limitations</h1>● No GGUF support The pack loads standard safetensors checkpoints; GGUF quantized weights are not supported, so GGUF-based quantized / CPU-offload paths cannot be used.● Streaming loader is slower Reading weights from disk on demand makes initial load and generation startup slower than eagerly loading the full model into RAM. It trades speed for a lower peak memory footprint on low-RAM machines.

Files with SHA256: c5384154108b7c0215d9a5b6c5017987206cde129d50b19748b0d33a367ba839