DEV Community

AI-KSK
AI-KSK

Posted on

Bernini-R: Reference Video Editing That Handles Subject Swap, Background Swap and Plain V2V

#comfyui #ai #tutorial

Bernini-R is a Wan2.2-based video model that renders text, images, videos and multiple references into a video. In practice that means one graph family covers subject replacement, background replacement, reference-video-conditioned editing, image reference, V2V and even text-to-image. Nine graphs ship. The author's summary is that reference conditioning is where it shines, and plain image-to-video is not.

Evidence note: nine workflow JSONs ship with this episode; model files, the two-stage KSamplerAdvanced values and every LoRA weight below are read from them. Rankings and task-type semantics are the author's, as reported in the video.

When to use it

  • You need to swap a person or a background in existing footage without rotoscoping.
  • You want one model family for V2V edits, reference edits and ordinary generation.
  • You are already on Wan2.2 and want to reuse the LoRA stack.

What is different in this workflow

  • One model, five task types: RV2V (video plus reference image), VRC2V (video plus reference video), V2V (video only), R2V (reference image to video), plus T2V/T2I.
  • No mask for subject replacement - the author says you just describe the swap and let the model do it.
  • An LLM node builds the prompt, because Bernini's text encoder is the old one and raw prompts behave oddly; the author uses Gemini 3.1 Pro Preview for the multimodal (image + video) prompt.
  • Two-stage sampling on high-noise and low-noise UNETs, which is the Wan2.2 pattern.
  • Wan2.2 LoRA stack carried over wholesale, including a rank-128 reward LoRA used at 3.0 on the high-noise branch.

Step by step

  1. Subject replacement: load the source video and the reference image, describe the swap in the LLM node, push its output into the positive prompt, set task type RV2V. 0.8 MP is 720p; drop to 0.6 for speed.
  2. Background replacement: load video plus background image, task type RV2V, short prompt, run.
  3. Reference-video condition: feed a reference video instead of an image; the task becomes VRC2V.
  4. Plain V2V: load only the target video, set task type V2V, describe the edit (the author's example: put the woman in a Lolita dress).
  5. R2V: one reference image plus a simple prompt.
  6. Multi-reference: 3 images through a Batch Images node into reference image, task type R2V - the author insists multi-image must go into reference image or it may not take effect.
  7. T2V / T2I: switch the task type, set length to 1 frame for T2I, more frames for video.
  8. Use the LLM node for anything with a video input; the author calls Bernini's own prompt behaviour very confusing.

Parameter basis

  • Models (all nine graphs): UNETLoader -> Bernini_HIGH_fp8_e4m3fn_scaled.safetensors and Bernini_LOW_fp8_e4m3fn_scaled.safetensors; CLIPLoader -> umt5_xxl_fp8_e4m3fn_scaled.safetensors (type wan); VAELoader -> wan_2.1_vae_Comfy-Org.safetensors.
  • Sampling: two KSamplerAdvanced at 14 steps, CFG 1, euler, simple - first enable/0-10, second disable/10-10000.
  • LoRAs: Wan22_T2V_UnifiedReward-Flex_LoRA_HIGH_rank64_bf16 and ..._LOW_rank64_bf16 at 1.0; wan2.2_t2v_A14b_high_noise_lora_rank64_lightx2v_4step_v1217 and wan2.2_t2v_A14b_low_noise_lora_rank64_lightx2v_4step_v1217 at 1.0; lightx2v_T2V_14B_cfg_step_distill_v2_lora_rank128_bf16 at 3.0 (high noise) and 1.0 (low noise); some graphs also load wan2.2_t2v_lightx2v_4steps_lora_v1.1_low_noise at 1.5 or 3.0.
  • Node counts: subject swap 47, background swap 48, V2V 39, single reference 39, 3-reference 44, image edit 37, I2V 39, T2V 35, T2I 32.

Limits

  • Reference conditioning is the strength, plain I2V is not: the author says the same image through LTX2.3 looked better than through Bernini's image-to-video.
  • The text encoder is unchanged from the original Wan stack, and the community believes one encoder was not open-sourced - the author blames this for needing an LLM to write prompts.
  • Multi-image must go through the reference image input or it silently may not apply.
  • Background swap and subject swap still occasionally need prompt iteration, which is why the LLM node matters.
  • Watermark removal is possible but the author is not ready to claim it beats the LTX2.3 route.

Episode resources

Top comments (0)