🎞️🖼️AGSoft MiniMaxH3 Images Slice

MiniMax H3 only generates one clip at a time - 4 to 15 seconds, however you configure it - so anything longer is a chain of scenes, and each new scene wants to know what the last one looked like. The way ComfyUI's native H3 support expresses that is first_frame (a still the next scene starts from) plus a few context_images as references. This node makes those two things out of frames you already decoded. If your workflow already runs a VAE decode for a preview or a save node, grabbing them here costs nothing: no second decode, no extra VRAM, just a tensor slice.

The name is doing honest work: it slices images, not latents. Feed it the decoded frames of a finished scene (images, plain 4D [F, H, W, C] or 5D batched [B, F, H, W, C] - both are accepted and normalized internally) and it returns:

• first_frame - the very last frame, always as 4D [1, H, W, C] IMAGE. That shape matters: the H3 image-to-video node's Qwen3VL reference encoder throws "too many values to unpack" on a 5D batch, which is exactly why the author pins it to 4D. Wire this into the next scene's first_frame input.
• context_images - the last context_frames frames (default 5), for the reference / motion-control inputs.

The only input you'll actually touch is context_frames (1–32). Fewer frames means less VRAM and fewer vision tokens fed downstream, so 1–3 is the weak-GPU sweet spot; bump it to 5–8 when you're chasing hard character and scene consistency and have the headroom. If your clip is shorter than what you asked for, the node logs a bilingual "context reduced to N" notice rather than erroring - it just gives you everything it has.

A couple of notes before you wire it in. This is a utility node in a huge one-author pack (comfyui-AGSoft is a grab bag of ~150 text, loop, video and H3 helpers, docs written in Russian with English alongside, and the author's documentation lives on a Telegram channel). It needs no model downloads and no API key - but the H3 model and VAE you're slicing are your own problem to have running first, and the H3 weights carry a community licence that geofences local use out of the US, EU, UK and Korea.

Install: ComfyUI Manager → search comfyui-AGSoft → install, or the manual route:

cd ComfyUI/custom_nodes
git clone https://github.com/Art-xmaster/comfyui-AGSoft

Restart ComfyUI. The pack's own dependency list (numpy, opencv-python, translators) is for its other nodes - you can install it clean and ignore the rest.

Use this node whenever your workflow already decodes the scene somewhere. If it doesn't - if you're holding only the latent - the sibling AGSoft MiniMaxH3 Latent Slice decodes for you instead. Choosing wrong isn't fatal; it just means paying for a decode you didn't need.