🎞️🧊AGSoft MiniMaxH3 Latent Slice

MiniMax H3 hands you ~4–15 seconds per generation, so a longer video is a chain of scenes, and every scene after the first needs to know what came before. H3's own continuation inputs are a first_frame still plus reference context_images. This node pulls both out of a finished scene's latent, doing the VAE decode itself so you don't have to restructure a latent-only workflow around a decode you never wanted.

Hook it up like this: latent_input takes the completed scene latent straight off the sampler - H3 latents are NestedTensors bundling the 5D video part and an audio part, and the node automatically picks the video part and ignores the audio. vae is the MiniMax H3 video VAE (minimax_h3_video_vae_fp16), the same one you'd use for a normal decode. Then it decodes just that latent and slices:

• first_frame - the final frame as 4D [1, H, W, C] IMAGE, shaped deliberately so the next scene's H3 image-to-video node (whose Qwen3VL reference encoder balks at 5D batches) accepts it cleanly.
• context_images - the last context_frames frames (1–32, default 5) for references and motion control.

Those two outputs are the whole point: wire first_frame into the next scene's first_frame input and you get the same character, pose, scene and camera flowing across the cut. That's how multi-scene H3 storytelling is assembled in practice - generate scene A, slice its tail, generate scene B anchored to it, repeat.

The only knob you'll regularly touch is context_frames. The trade-off is spelled out in the tooltip and it's the honest one: more frames = better motion consistency but more Qwen3VL tokens and VRAM downstream. On a weak GPU keep it to 1–3; go 5–8 when you need hard character consistency and have the memory. If the clip is shorter than requested it degrades gracefully and logs a notice instead of failing.

Now the timing question, because it's the whole point of this node's existence: decoding a latent is not free. The family here is split by it - if your workflow already decodes the scene somewhere (a preview, AGSoft Video Save, an upscaler pass), reach for the sibling AGSoft MiniMaxH3 Images Slice, which cuts the same context out of pixels you've already paid for. Use Latent Slice when nothing downstream needs the decoded frames and you'd otherwise decode only for context - or when you're deliberately keeping the pipeline latent until the very end.

One thing to know about the pack: comfyui-AGSoft is a big single-author collection (~150 utility nodes, Russian-first docs, a Telegram channel for guides), and it has essentially no English community discussion yet. These MiniMaxH3 nodes sit in the AGSoft/MiniMaxH3 category and assume you already have H3 running - no extra model downloads come from the pack itself. Install via ComfyUI Manager (search comfyui-AGSoft) or:

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

Then restart. Just remember the H3 weights themselves carry a community licence that excludes the US, EU, UK and Korea from local use - that's a hurdle before any of these nodes enter the picture.