<h1 id="comfyui-bruxos-do-vfx">ComfyUI Bruxos do VFX</h1><p><a target="_new" rel="ugc" href="https://github.com/NyckM/Bruxos-do-VFX-Nodes/tree/main">https://github.com/NyckM/Bruxos-do-VFX-Nodes/tree/main</a></p><p>Custom nodes for video VFX with Wan 2.2 / Bernini: object and people removal, batch upscaling, face swap, A/B comparison, and production-ready utilities — all designed for real-world VFX workflows.</p><p>&lt;img width="832" height="480" alt="Workflow for object and people removal" src="https://github.com/user-attachments/assets/6745a8ce-00ce-4915-a60d-ed1354099311" /&gt;</p><h2 id="why-this-exists">Why this exists</h2><p>Developed at <strong>Bruxos do VFX</strong> during the production of two feature films — <strong>Dr. Monstro</strong> (Marcos Jorge) and <strong>Alice Júnior 2</strong> (Gil Baroni). The complexity of compositing and scene integration created the need for a dedicated object-removal workflow built on top of Bernini/Wan.</p><p>Current performance averages <strong>96 frames in 204 seconds</strong>.</p><p><a target="_new" rel="ugc" href="https://github.com/user-attachments/assets/1575f97f-34b9-492a-accb-818e97a6cbde">https://github.com/user-attachments/assets/1575f97f-34b9-492a-accb-818e97a6cbde</a></p><p><a target="_new" rel="ugc" href="https://github.com/user-attachments/assets/d1e47486-ac3b-4030-be42-56a0f16b0128">https://github.com/user-attachments/assets/d1e47486-ac3b-4030-be42-56a0f16b0128</a></p><hr /><h1 id="installation">Installation</h1><h2 id="easy-(recommended)">Easy (recommended)</h2><p>Run the included <code>.bat</code> installer from your ComfyUI installation folder. It will automatically install the node pack and download all required models.</p><p><em>(RunPod version is still under development.)</em></p><h2 id="manual">Manual</h2><p>Copy the package into:</p><pre><code></code></pre><pre><code>ComfyUI/custom_nodes/ComfyUI-Bruxos-do-VFX</code></pre><p>Restart ComfyUI.</p><h2 id="required-models">Required Models</h2><p>ModelDownloadDestinationUNET GGUF (high/low)<a target="_new" rel="ugc" href="https://huggingface.co/neuregex/Bernini-R-GGUF/tree/main">https://huggingface.co/neuregex/Bernini-R-GGUF/tree/main</a><code>ComfyUI/models/unet</code>Distill LoRA (4 steps)<a target="_new" rel="ugc" href="https://huggingface.co/Cyph3r/lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank64_bf16">https://huggingface.co/Cyph3r/lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank64_bf16</a><code>ComfyUI/models/loras</code>Text Encoder (umt5)<a target="_new" rel="ugc" href="https://huggingface.co/Osrivers/umt5_xxl_fp8_e4m3fn_scaled.safetensors">https://huggingface.co/Osrivers/umt5_xxl_fp8_e4m3fn_scaled.safetensors</a><code>ComfyUI/models/text_encoders</code>VAE<a target="_new" rel="ugc" href="https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Wan2_1_VAE_bf16.safetensors">https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Wan2_1_VAE_bf16.safetensors</a><code>ComfyUI/models/vae</code></p><p><strong>Important:</strong> Download the models using the Hugging Face CLI or a download manager. Browser downloads may produce incomplete or corrupted files.</p><hr /><h1 id="nodes">Nodes</h1><p>&lt;img width="226" height="557" alt="Bernini Long Condition" src="https://github.com/user-attachments/assets/934fd9b3-087f-47db-aeb2-6ca01967c556" /&gt;</p><h2 id="bernini-generation">Bernini / Generation</h2><ul><li><p><strong>Bernini Infinity</strong> — Main renderer for videos longer than the standard 81-frame limit without requiring a custom sampler. It injects <code>context_latents</code> per chunk (with optional <code>tail_memory</code>) instead of relying on a single conditioning, taking advantage of Bernini/Wan's native architecture.</p></li><li><p><strong>Bernini Region Mask</strong> — Converts colored masks (SAM2/SAM3/Scail2Color) into black-and-white masks with invert, grow, and blur controls.</p></li><li><p><strong>Bernini Long</strong> <em>(Conditioning / ChunkSelect / VideoMerge / AppendVideoChunk / EmptyVideoChunks / Info)</em> — Helper nodes for long video generation.</p></li><li><p><strong>FaceStitchUpscale</strong> — Blends the upscaled face back into the video using the <code>face_bboxes</code> generated by Pose and Face Detection.</p></li><li><p><strong>SAM3 Point Editor</strong> — Green click = positive point, purple click = negative point directly on the frame, providing much more stable tracking than text prompts alone.</p></li></ul><h2 id="video">Video</h2><ul><li><p><strong>Load Video / Save Video</strong> — Native VIDEO (ComfyUI 2.0) equivalents to Video Helper Suite, including preview trimming (skip, cap, nth frame, force rate) and extended export controls for codec and CRF.</p></li><li><p><strong>Compare Videos A/B</strong> — Built-in comparison player with wipe, side-by-side, difference, and toggle modes, allowing before/after inspection without leaving ComfyUI.</p></li><li><p><strong>Mask BBox Preview</strong> — Displays the bounding box that <code>bbox</code> mode would crop before rendering.</p></li></ul><h2 id="upscaling">Upscaling</h2><ul><li><p><strong>Upscale Config / Batch Blend</strong> — Super-nodes that replace the original Settings and Blend Frames subgraphs.</p></li></ul><h2 id="utilities">Utilities</h2><ul><li><p>Grow + Blur Mask</p></li><li><p>Block Mask</p></li><li><p>Smart Sharpen</p></li><li><p>Text Utilities / Show Text</p></li><li><p>Seed Generator</p></li><li><p>Load Images From Folder</p></li><li><p>Video Info</p></li><li><p>Wan All-in-One Loader</p></li><li><p>Qwen-VL Caption</p></li><li><p>Prompt Guide (official Bernini presets)</p></li></ul><hr /><h1 id="sequential-vs-context-window">Sequential vs Context Window</h1><p><strong>SequentialContext Window</strong>ProcessingProcesses chunks sequentially, advancing <code>chunk_size - overlap</code> frames each stepProcesses the full video using a sliding context windowVRAM UsageLowerSlightly higher<code>mask_mode = bbox</code>❌ Falls back to <code>inpaint</code>✅ SupportedTemporal ConsistencyGood (via <code>tail_memory</code>)Better (native)Recommended UseVery long videos or limited VRAMVideos that fit in a single generation or when using <code>bbox</code> mode</p><p>⚠️ Using a very small <code>chunk_size</code> together with a large <code>overlap</code> dramatically increases the number of passes (for example: <code>chunk=17</code>, <code>overlap=16</code> results in <strong>61 passes</strong>). Larger chunks with smaller overlaps are generally recommended.</p><h3 id="mask-modes">Mask Modes</h3><ul><li><p><strong>off</strong> — Regenerates the entire frame.</p></li><li><p><strong>inpaint</strong> — Only edits the masked region while preserving the original image elsewhere.</p></li><li><p><strong>bbox</strong> — Crops the masked region and generates it at a smaller resolution for significantly lower VRAM usage. <em>(Available only in Context Window mode.)</em></p></li></ul><hr /><h1 id="changelog-(major-milestones)">Changelog (Major Milestones)</h1><h3 id="v0.2">v0.2</h3><ul><li><p>Automatic <strong>4n+1 frame correction</strong> (mirrored padding + final trimming)</p></li><li><p>Added <code>mask_mode</code> (<code>off</code>, <code>inpaint</code>, <code>bbox</code>)</p></li><li><p>Added <code>mask_grow</code></p></li><li><p>Added <code>mask_blur</code></p></li></ul><h3 id="v0.3">v0.3</h3><ul><li><p>Added <strong>FaceStitchUpscale</strong></p></li></ul><h3 id="v0.5">v0.5</h3><ul><li><p>Added native <strong>VIDEO</strong> Load/Save nodes (ComfyUI 2.0)</p></li></ul><h3 id="v0.6-0.9">v0.6–0.9</h3><ul><li><p>Complete suite of native utility nodes (reducing third-party dependencies)</p></li><li><p>Compare Videos A/B</p></li><li><p>Mask BBox Preview</p></li><li><p>Upscale Config / Batch Blend</p></li><li><p>Server-side preview trimming in Load Video</p></li></ul><h3 id="v0.10">v0.10</h3><ul><li><p>Added <strong>SAM3 Point Editor</strong> (green positive / purple negative points) for significantly more stable object tracking.</p></li></ul><hr /><h1 id="why-isn't-there-a-&quot;bernini-long-sampler&quot;">Why isn't there a "Bernini Long Sampler"?</h1><p>Because Bernini already passes temporal information through conditioning, and the Wan patch already accepts <code>context_latents</code> as a list. The important part is generating the correct conditionings:</p><pre><code></code></pre><pre><code>context_latents = [
    encoded_chunk,
    tail_latent,
]</code></pre><p>This approach leverages the native Bernini architecture instead of duplicating an entire sampler, making the package easier to maintain and fully compatible with existing Bernini workflows.</p><p>An automatic chunk executor may be added in the future, but it would depend on the exact node names and class implementations installed on each user's system.</p><hr /><h1 id="acknowledgements">🙏 Acknowledgements</h1><p>Built on top of Kijai's work and the Bernini models.</p><p>Many thanks to the original authors and the open-source community.</p><h1 id="license">📄 License</h1><p>Apache License 2.0.</p><p></p>

Files with SHA256: f81f5f414bdd6d5a1649e65ac5271a31effda2975c058a1b3bc240474894ae20