AICG-渲染器（高级）

The appeal is the wire count

A stock H3 render is five nodes before you've decoded anything: noise, scheduler, sampler select, basic guider, custom-sampler-advanced. Then two decoders, because H3 emits picture and audio from the same latent, and then a CreateVideo to glue them into one file with a frame rate. Fourteen-ish nodes, six of which you set once and never touch again except when something drifts.

AICG-渲染器（高级） (AICG3D_H3RenderAdvanced) collapses all of that into one box with one input and one output. Feed it the h3_context from an H3 context node, get a VIDEO out, wire that to SaveVideo. If your workflow is one clip and you're not doing latent surgery in the middle, this is the node you want - the pack's own 13-series workflows lean on it for exactly this reason.

What happens inside

Reading the source, the render function does three things in order.

Sampling. It runs the same internal path as AICG-采样器（高级）: build noise from noise_seed, compute sigmas from scheduler + steps + denoise, and sample with the conditioning and latent that live inside the context. If you connect the optional sampler and sigmas inputs - an external sampler pair like a DMAD four-step setup - those take over, and the tooltips are explicit that sampler overrides sampler_name while sigmas overrides scheduler, steps and the denoise truncation.

Decoding. It tries the official joint AV decode first, splitting the combined latent into video and audio and running each through its own VAE. If that helper isn't importable it falls back to a plain VAEDecode plus vae_decode_audio - same result, slightly different path, and it means a half-broken install still gets you a picture out rather than a stack trace.

Composing. Frames and audio go into a VIDEO at the context's fps. One wrinkle worth knowing: if the context carries a locked driving audio track (Digital Human mode), that track is used instead of the generated one. Your avatar says the line you gave it, not the line the model invented.

The settings that actually matter

• noise_seed - 64-bit. Move it until the shot works.
• sampler_name / scheduler / steps / denoise - the standard quartet, defaulting to 20 steps at full denoise. On H3 those defaults are fine; if you've come from SDXL, resist the instinct to pile on steps, and if you've bolted on a distilled LoRA, drop the steps to whatever that LoRA documents.
• cleanup_after_run - the one people don't read. 卸载模型 (the default) unloads the model and empties the cache after the node runs: lowest memory, slowest next run. 释放缓存 returns cached VRAM to the driver but keeps the model resident - noticeably faster on a second render. 不处理 leaves ComfyUI alone. On a 24GB card running a 33B model, the default is the right call; if you're doing three takes in a row on 48GB, switch it to 释放缓存.
• Optional model - a patch chain. Connect it and the renderer samples your model instead of the one in the context, which is how you apply a LoRA or a sampling patch without rebuilding the context.

One output: video (VIDEO), straight into SaveVideo or a combine node.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/JGRFW/comfyui-AICG3D

Restart, then Ctrl+F5 - the frontend JS is cached hard enough that a soft reload can leave you looking at stale panels. Manager route: search the Registry for AICG3D, 短剧工作台 or MiniMax H3 (publishes as aicg3d-h3-suite).

Dependencies genuinely are light: requirements.txt contains exactly imageio-ffmpeg. If ComfyUI's Python can't find it:

python -m pip install imageio-ffmpeg

Run that in ComfyUI's environment, not your system Python. The model side is the real prerequisite - FL2VA (and optionally REF2VA) transformers, the text encoder, the video VAE and the audio VAE, all present in the folders the loader scans.

Where it bites

Memory, honestly. H3 is a 33B model; people asking whether it runs on a 3060 got the answer "slowly" on launch day, and that has not changed. This node's default cleanup is aggressive for a reason. If you OOM at the decode rather than the sample, lower resolution or frames before you touch steps.

denoise less than 1 with an external SIGMAS connected. The node logs that it's ignoring your denoise because the external schedule wins. That's correct behaviour, but if you were expecting a partial re-render, you'll get a full one.

Silent output. You skipped a VAE somewhere upstream, or the context was built in Digital Human mode with a driving track that's shorter than the clip.

Nothing renders at all. Check the context came from a whole-clip H3 node. Contexts built by the Context Segments path are multi-shot plans, and the single-pass renderers reject them by design rather than silently rendering shot one.