Anime2Real Bernini Conditioning

This is the simplest of the pack's Bernini-style conditioning nodes, and the pack's own description explains exactly why: it's "pure v2v anime->real conditioning" with no reference image involved at all. If BFSHeadSwapBerniniConditioning is about splicing a new identity onto a video, this one is purely a style transform - anime source in, photorealistic output, same motion and layout throughout.

How it's wired internally, and why that matters for CFG

The node description spells out the conditioning layout precisely, which is worth repeating because it changes how you should think about sampler settings: guide_video gets tagged as source_id 1 (the anime canvas) while the output occupies source_id 0 (photoreal). There's no reference image, so nothing gets the special positive-only treatment BFSHeadSwapBerniniConditioning uses for identity amplification. Instead, guide_video is placed on both positive and negative conditioning - meaning CFG in this setup only amplifies your text prompt (the photorealistic style you're asking for), not any injected reference, since there isn't one. The practical takeaway: don't crank CFG expecting stronger anime-to-real transformation the way you might with the head-swap node's amplify_reference on. A moderate CFG is what the author recommends; the guide video itself is what keeps motion and layout locked in place regardless of CFG.

The node's own docs also flag a specific incompatibility worth remembering: don't run the head-swap node against an anime2real LoRA. Head-swap conditioning injects a source_id 2 token that this LoRA family was never trained to see, so the two aren't interchangeable even though they share a conditioning family.

Inputs and outputs

Only four required fields: positive, negative, vae, and guide_video (your anime source - output size matches it). length sets frame count, snapped to 4k+1, trained at 41 frames - notably shorter than the head-swap node's 73-frame default, so don't assume you can reuse settings between the two without adjusting.

Outputs are positive, negative, latent, and debug - the standard shape for this pack's Bernini-conditioning nodes, ready to feed straight into your LTXV sampler.

Setting it up

Install through ComfyUI Manager (search "ComfyUI-BFSNodes") or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/alisson-anjos/ComfyUI-BFSNodes.git
cd ComfyUI-BFSNodes
pip install -r requirements.txt

Restart ComfyUI. As with the rest of this conditioning family, you need an anime2real LoRA trained specifically for this source_id 0/source_id 1 layout, loaded onto your LTXV model with a standard LoRA loader before your sampler - this node only builds the conditioning that tells the model which source is which; the LoRA is what taught the model to actually perform the transform.

What tends to go sideways

Because this node has no strength or amplification knob of its own, most problems trace back to the sampler or the source footage rather than the node itself. If the output looks like a color-graded anime clip rather than a genuine photoreal transform, that's a LoRA-strength or CFG issue on your end, not something this node can fix - there's no reference to amplify here. If motion looks broken or frames feel misaligned, double-check length against what your source clip actually contains; 41 frames snapped to the nearest 4k+1 is a specific number, and feeding it a mismatched clip length is an easy way to get artifacts that look like a model problem but aren't. And if you're tempted to add identity control on top (say, keeping a specific character consistent across an anime-to-real pass), that's a job for a different node in this pack, not a setting to bolt onto this one - it wasn't built for it.