kubakub region video sampler

Video diffusion and architecture have a compatibility problem: the model wants to move the whole frame. Camera drift, a slight pan, the whole scene breathing - that's the aesthetic. Projection mapping is the opposite. The building must not move, because the pixels are landing on real stone and any drift is immediately visible as content sliding off a window frame.

kubakub region video sampler (ltx) exists for exactly that. It animates only the regions your plan says animate = on, on LTX 2.x, and pastes the result back into a still frame after still frame - so everything that isn't an animated region is not merely stable, it's bit-identical to the input.

How it works

The plan decides. Regions sharing the same video_prompt, t_start, t_end and motion are grouped into one clip, because there's no point running the model twice for two windows that are doing the same thing. Each clip runs on its own crop, sized to the LTX grid, with the rest of the crop held by a noise mask - so the model sees the surrounding facade but isn't allowed to change it. The first frame is your still. Then the result is pasted back into the untouched still, frame by frame, with a feather_px soft edge. Beyond the feather, nothing was ever touched.

The whole illusion rests on that paste. The moss growing over a window is generated; the mortar around it is literally the same bytes as the input image.

The inputs

model is LTX 2.5 22B distilled, clip is the Gemma text encoder loaded as type ltxv, and you need both vae and audio_vae - the model is audio-video, and without an audio input the audio just stays silent. plan from kubakub region plan, with keys animate, video_prompt, t_start, t_end and motion. And image is the still (your frame-in-frame + region sampler result, or just a photo).

Then:

• frames - length, rounded up to 8k+1 because that's what LTX wants. Ignored when background is connected, since then the video is as long as the background.
• fps - the plan's t_start/t_end are in seconds, so this is what makes them mean anything.
• stage1_mp - megapixels per clip in the first stage, doubled per side by the upscaler. 0.2 by default.
• context_px - how much still each clip sees around its regions. 64 by default; raise it if a clip's content looks unconnected to its surroundings.
• output_scale - 0.5 by default, and the tooltip is brutally honest about why: a 4K batch of 121 frames needs 12 GB of RAM just to hold the output. Use save_folder for full-size PNGs instead.
• stage2_max_mp - the cap for the sharper second stage, and the pack's own note is the most useful line in the node: 0.6 MP × 121 frames is safe on a 12 GB card, 1.8 MP crashed it. Take that seriously.

Optional but worth knowing: upscale_model (the LTX latent spatial upscaler ×2) makes the second stage actually sharper. only narrows to a selection in plan syntax. negative is one negative prompt for all clips, and it ships with a good one - camera movement, zoom, pan, shaky, morphing facade, blurry, low quality, text, watermark - which tells you the failure mode this node was built to fight. loop makes each clip return to the still so the video repeats, with calmer motion near the end because it has to come back. audio holds a sound the clips are sampled with; with an audio-reactive LoRA and its trigger words in video_prompt, the regions move with the music. enhance_prompt + enhance_clip expand each clip's prompt with a writing text encoder - the Gemma LTX int8 encoder can't generate text, so you need a separate one like a Qwen.

Outputs

frames, report (clips, frames, stages, time, one line per clip), fps (passed through for Create Video), audio (your sound cut or padded to the length of the frames, silence if none connected), and mask (the animated regions at the frame size). Note the audio caveat from the tooltip: the frames carry no sound - add the same audio in Create Video.

Install, and the honest VRAM conversation

cd ComfyUI/custom_nodes
git clone https://github.com/kubakubkub/ComfyUI_KubaNodes
python -m pip install -r ComfyUI_KubaNodes/requirements.txt

Manager → kubakub nodes. ComfyUI 0.37.0+. No Blender, but you need an LTX 2.x video model, which is where the real cost is.

The pack says it was tested on a 12 GB GPU, and it means 12 GB with quantised weights. The wider community reality for LTX-2.3 22B is that quantised GGUF workflows are what make 12 GB work at all, and that 64 GB of system RAM is the comfortable pairing. The relevant point for this node: it eases the pressure enormously, because each clip is a small crop rather than a full-frame generation. You're not running 4K video generation - you're running a 0.2 MP crop of a window, several times.

Clips are cached individually and models load once for all clips, so the second pass at the same settings is dramatically faster than the first - and while the director is holding or nothing is animated, the models aren't loaded at all.

Two things to expect

Motion has to be describable. "The window becomes an aquarium" runs. "Subtle atmospheric change" doesn't, on any video model. Short vague prompts are the most common reason a clip comes back nearly still.

only uses the plan's selector syntax, and 0.3.0 fixed it so group: and tag: work and the report says when a selector matched nothing. Before that, a group: selection silently rendered no clip at all - worth knowing if you're reading an old issue thread.