AICG-渲染器（高级·加速）

The difference is one socket

AICG-渲染器（高级·加速） is AICG-渲染器（高级） with a single structural change, and if you don't notice it you'll spend ten minutes wondering why the node won't queue. In the plain version, model is optional - the renderer happily uses whatever model the h3_context is carrying. Here, model is a required input, and it comes first in the socket order, above h3_context.

That's it. Same sampling path, same joint audio/video decode, same VIDEO output, same noise_seed / sampler_name / scheduler / steps / denoise / cleanup_after_run widgets, same optional sampler and sigmas for external schedules. The source literally subclasses the other node and reorders the inputs.

So why does the pack ship both?

Why a required model wire is the better default

Because it makes the patch chain visible. H3's "acceleration" story is LoRAs and sampling patches - a four-step distilled LoRA, a sigma-shift node, whatever model surgery your workflow needs - and all of that happens on a MODEL between the loader and the sampler. In the plain renderer that wire is optional, so it's easy to leave it dangling and sample the base weights while your carefully loaded LoRA sits in the graph doing nothing. The model still comes from the context, the render still works, the video still looks plausible, and you have no idea your patch was ignored.

The WithModel variant takes that failure mode off the table. No model wire, no run. Two tracks on the canvas: model on top, context below.

How to wire it

Load with the pack's loader (which emits both h3_bundle and model), or build a bundle with AICG3D_H3ModelAdapter and patch the model yourself. Then:

MODEL  -> [LoRA loader] -> [ModelSampling / shift patch] -> model  (this node)
h3_context                                                 -> h3_context
video                                                      -> SaveVideo

If you have no patches at all, connect the loader's model straight in. The node doesn't care whether the chain has anything in it - it just wants the wire to exist.

Settings, briefly

Everything you'd set on the plain renderer applies here:

• noise_seed - 64-bit, reroll until the shot is right.
• sampler_name, scheduler, steps, denoise - 20 steps at full denoise is a sane default for H3 at native resolution.
• cleanup_after_run - defaults to unloading the model and emptying the cache after the run. Lowest footprint, slowest second take. Switch to 释放缓存 if you're iterating on one shot with VRAM to spare.
• sampler / sigmas (optional) - external schedule overrides, tooltip-documented as overriding sampler_name and the scheduler/steps/denoise trio respectively.

Output is the single VIDEO.

Install

Manager: search AICG3D or 短剧工作台. Manual:

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

Restart ComfyUI and hard-reload the browser (Ctrl+F5) - this pack's UI assets are versioned precisely because stale JS causes ghost bugs. The only Python dependency in requirements.txt is imageio-ffmpeg.

You'll need H3 weights in place before any of this is meaningful: FL2VA/REF2VA transformers in models/diffusion_models, the text encoder in models/text_encoders, minimax_h3_video_vae_* and minimax_h3_audio_vae_fp32 in models/vae. Note that the H3 Community License carves the US, EU, UK and South Korea out of its territory - worth reading before you build a pipeline on it commercially.

Traps

"Required input missing: model." The whole point of the node. If you migrated an existing graph from the plain renderer by swapping the class, the wire didn't come with it.

Patching after the context. If you also built the context through a path that applied its own acceleration patches, you now have two model states in play: the context's and the one on this wire. The explicit wire wins, which is the behaviour you want, but it means the context's patched model is idle and its VRAM is wasted. Pick one place to patch.

Expecting it to be faster. "加速" in the display name isn't a performance claim from this node - it's the variant that makes an acceleration chain practical. The speed comes from the LoRA and the schedule you attach, not from the node class.