Bernini-R Conditioning

If you've used a stock ComfyUI graph, the empty latent, the VAE encode of your source, and the mask handling were three separate chores. BerniniR_Conditioning folds all of them into one node, and it's the closest thing this pack has to a workflow in a box. It creates your starting latent, encodes whatever visuals you give it into context_latents that the model attends to during denoising, and wires it all onto your conditioning. It's also where "Bernini-R does editing without a mask pipeline" stops being strictly true - the mask input is here, optional, and very good when you connect it.

How it works

You feed it your positive/negative conditioning (from BerniniR_PromptEmbedding), your VAE, and a width/height/length. It builds a zero latent and, if you connect any visual input, VAE-encodes it into context latents: source_video for video-to-video, reference_video for style/motion guidance, reference_images for subject/style conditioning (the node autogrows up to eight reference slots). Because encoding a long source video is exactly the kind of thing that OOMs small GPUs, encoding happens in chunks (chunk_frames, default 16 pixel frames) with overlap so the VAE's 3D convolutions don't leave temporal seams.

The inputs that matter

• width / height / length - output dimensions and total pixel frames. The README's starter values: 832×480, length 81 (~5 seconds at 16 fps).
• positive / negative / vae - from the prompt planner and VAE loader.
• source_video / reference_video / reference_images - the in-context visuals. Which ones you connect is what actually determines the task (prompt-only = text-to-video; prompt + reference image = I2V; prompt + source video = V2V editing).
• batch_size - >1 creates multiple latents for parallel generation in one run. Fun, but watch VRAM like a hawk.

The mask pair deserves its own paragraph:

• mask - from a segmenter like SAM2. White (1) = regenerate, black (0) = keep source, same convention as ComfyUI's denoise_mask. Connecting it enables differential diffusion. You connect it without inverting.
• mask_mode - anneal (default) = soft background anchoring with natural boundaries; freeze = pixel-level freeze where the background stays completely still. Start with anneal; freeze is for when the background must not move a single pixel.

Outputs and wiring

Three: positive and negative (CONDITIONING) - now carrying the context latents - plus latent (LATENT). All three go to BerniniR_KSampler (or the DualExpertSampler).

Where people get burned

The two classic issues both live in this node. OOM when encoding a long source video → drop chunk_frames lower (16 is already low, but a long clip at high res still adds up) and consider force_offload in the prompt planner. Mask edits that look wrong → check you didn't invert the mask; it must be white-where-you-edit, straight from SAM2, no inversion. And remember the model-level truth that makes all editing here heavy: the source video becomes part of the model input, so an edit costs roughly double the compute of a plain Wan 2.2 generation. That's not your workflow's fault. It's the model.