Silver H3 Batch (N videos / encode once / sample once / decode once)

Render five clips the stock way and you load the text encoder five times, the DiT five times, the VAE five times. Then the clips come back and you notice the generation wasn't the slow part - the shuffling of tens of gigabytes in and out of VRAM was. SilverH3Batch exists to fix exactly that and nothing else.

It renders N independent MiniMax-H3 videos (video plus the model's native audio, 24 fps) from N prompts, and it keeps the discipline the stock queue can't: the text encoder, the DiT and the VAE each load once for the whole batch. A twelve-clip run costs about two model swaps instead of roughly 3 × 12. For a 33B model with a big text encoder, that's not a small win.

How it works

The node is built as three phases, which you don't drive - it just stops interleaving them per clip:

• Phase A, ENCODE. Every prompt is text-encoded in one text-encoder session. Your keyframe images and character references are VAE-encoded here too, once, and reused.
• Phase B, SAMPLE. Every video is sampled back to back while the DiT stays resident. No mid-sampling keyframe decodes, which is why the "no mid-phase swaps" claim actually holds here.
• Phase C, DECODE. Every latent is decoded in one VAE session. Each clip is written straight to disk as it decodes and then freed, so at most one clip's frames are ever in RAM. That's how a twelve-clip batch fits on a consumer box instead of thrashing.

This node deliberately has no cross-clip continuity. Each prompt is its own short film, joined at hard cuts if you stitch them. If you want one continuous take across N shots, that's the sibling node, Silver H3 Staged Multishot.

The inputs that matter

prompts is a STRING_LIST, one prompt per video, and it's the input beginners fumble - it's a list socket, not a text box. Either wire it from one of the pack's two prompt helpers, or turn the widget into an input and feed it a list. overall_prompt is a plain string prepended to every prompt (overall + shot), which is where the scene-wide description belongs so you're not repeating it five times.

Then the usual shape knobs: width, height, frames_per_shot (243 default, about 10.1s, on H3's 17k+5 grid - the author's own note is that the trained range is roughly 124–362 and beyond that is RoPE-extrapolation territory), plus seed, steps, sampler_name and scheduler.

The interesting optional ones:

• start_image is an IMAGE batch, and image i opens clip i. Feed N+1 images and you get N clips chained frame-to-frame, because with combine_videos ON, clip i's last frame is automatically taken as image i+1's start - there is no separate end-image input.
• combine_videos (default OFF) affects saved files only. ON stitches everything into one master in a dated subfolder (clip_01.mp4, clip_02.mp4, … plus _master.mp4); OFF gives you one flat file per clip.
• char_refs_all rides every clip with stable <Picture N> labels; char_refs_index is positional, image i on clip i; char_audio is one voice shared by all.
• guide_audio is a spine locked per clip, with step_guide_audio deciding whether every clip reuses the head of the track or each advances to its own slice (auto-ON when combining).
• double_pass with double_pass_denoise (0.25) and double_pass_steps (8) is a Motion-Director-style refine over each full latent, and upscale_factor resamples the latent before decode - the suggested combo is 1.25x plus a refine pass. apply_diffaid applies the DiffAid sparse patch and hard-errors if that pack isn't installed, so leave it off unless you have it.

Outputs are audio (a list of audio dicts, one per rendered video), saved_videos (newline-separated paths) and videos_rendered.

Install

Manager search for "Silver Nodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/SilverAndJade/comfyui-silver-nodes

Restart ComfyUI. The pack's requirements.txt covers the older loader nodes (beautifulsoup4, moviepy, a pinned opencv-python==4.10.0.82 for NumPy compatibility) - the H3 nodes ride on ComfyUI's own PyAV and comfy_api.latest, so you need a reasonably current ComfyUI. And you need H3 weights: roughly 42.5GB, under the MiniMax H3 Community License, which excludes the US, EU, UK and South Korea from its applicable territory. If you're in one of those, the licence says the local weights aren't yours to run.

The repo ships SilverH3Batch_Turbo_3x10s.json and SilverFolderBatch_ToH3Batch.json - drag one in and read the wiring before building from scratch. Start with two or three prompts, not twelve.

Where people get burned

Keyframes and character refs are mutually exclusive. When refs are set, H3's cond_video_latents is replaced by the ref latents and the keyframes are wiped, so the node drops your refs with a warning if start_image is connected. An I2V clip cannot carry identity refs; a ref-driven clip cannot key its opening frame.

Don't trust a render count that looks short. videos_rendered is your prompts-list length, and blank entries are skipped - a stray empty line gives you fewer clips and no error to explain it.