CIVITAI / Workflows

Bernini-R Video-to-Video Editing Workflow

Watch the full video first if you want to understand how this Bernini-R V2V video editing workflow works in practice. The video shows how an existing source video can be edited through text instructions while preserving the original motion, camera structure, subject position, lighting, and scene rhythm. This ComfyUI workflow is designed for Bernini-R video-to-video editing. Its main purpose is to take a source video and apply a controlled visual edit without rebuilding the whole clip from scratch. Instead of using pure text-to-video generation, this workflow starts from real video frames, extracts the source video components, and then uses BerniniConditioning to guide the editing process around the original motion and composition. The workflow is built around the Bernini-R high-noise and low-noise model structure. It uses Bernini_HIGH_fp8_e4m3fn_scaled.safetensors and Bernini_LOW_fp8_e4m3fn_scaled.safetensors as the dual model branches. It also uses UMT5 XXL fp8 text encoding, Wan 2.1 VAE, LoadVideo, GetVideoComponents, BerniniConditioning, KSamplerAdvanced, VAEDecode, CreateVideo, SaveVideo, and PathchSageAttentionKJ. The model route includes LightX2V LoRA and UnifiedReward-Flex LoRA for both high-noise and low-noise stages, helping the workflow improve speed, stability, and final visual quality. The source video is the foundation of the workflow. LoadVideo imports the original clip, and GetVideoComponents separates the frame sequence, audio, and FPS. The extracted frames are then passed into BerniniConditioning as the source video condition. This means the original performance, body movement, camera angle, timing, and background relationship can remain stable while the edit instruction changes the visual result. The prompt system is also an important part of the graph. BerniniPromptEnhancer builds a Bernini-specific V2V instruction from a short user task. In the uploaded workflow, the example task is to make the woman in the video wear Lolita clothing. RHLLMChatNode then rewrites the task into a more detailed edit prompt. The output is cleaned through StringReplace nodes, removing the JSON wrapper before sending the final instruction into CLIPTextEncode. This allows the user to start with a simple edit request and let the workflow turn it into a more precise video editing prompt. The generation route uses BerniniConditioning with a vertical 480×848 se