MiniMax H3 Aicg Loader

Every H3 workflow in this pack begins here, and the node tells you why H3 isn't like loading a Flux checkpoint: it wants two transformer slots, an LLM-sized text encoder, and two VAEs. Miss any of it and nothing downstream will load.

MiniMax H3 Aicg Loader outputs h3_bundle (the thing every other node in the pack eats) and model (a plain MODEL, if you're hand-patching).

The two transformer slots

• fl2va_model - the first/last-frame backbone. Default none.
• ref2va_model - the multi-reference backbone used by reference-to-video and digital-human modes. The tooltip here is worth quoting because it's the single most useful paragraph in the pack: fused / hybrid / refdelta-style weights that already have reference capability built in also work in the FL2VA slot, so you can leave REF2VA empty. You only fill it if you specifically want the official standalone REF2VA weights - and those are not the turbo version, so give them enough steps.

Both slots list every weight on your machine, which is deliberate; pick wrong and you get a warning before sampling rather than a mysterious tensor mismatch. If both are none, the loader refuses: "Select at least one MiniMax H3 transformer."

The rest of the load

• text_encoder - H3 is LLM-encoded, not CLIP-encoded. The expected file is qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors, from models/text_encoders. A 32B vision-language encoder in NVFP4 AWQ: that's most of your download and a serious chunk of your VRAM. On a big Blackwell card NVFP4 is what makes local H3 plausible at all; on older hardware expect to feel it.
• video_vae - minimax_h3_video_vae_fp16.safetensors. Hosts a 3D latent. The tooltip warns you plainly that picking a normal image VAE like ae.safetensors errors at load.
• audio_vae - minimax_h3_audio_vae_fp32.safetensors, a 2D latent. The loader validates the roles of both VAEs, and will not let you put the same non-H3 VAE in both slots.
• lora_1 … lora_9 with lora_*_strength (−4 to 4). Yes, nine. They stack in order through LoraLoaderModelOnly, so slot order is application order. Speed distillers and style LoRAs live here; if you're running a 4-step distilled sampler, the matching LoRA usually belongs at strength 1.0 and the step count in the renderer has to agree.
• non_h3_policy - 自动回落 (ignore a non-H3 weight and use the other slot) or 按所选加载 (use exactly what you picked and hard-error if it isn't H3). The tooltip's example is an image model like Anima sitting in a slot. Leave it on the default unless you enjoy debugging.
• 注意力加速(Sage) - the one acceleration control the loader kept. MotionCache and its tuning moved to the dedicated MiniMax H3 Aicg 加速设置 node; the loader keeps exactly one widget so old workflows don't shift positions.

Outputs

h3_bundle → into MiniMax H3 Aicg, 上下文分段, 加速设置, or 模型中转. model → a MODEL you can send through a patch chain and then into a renderer's model input.

Installing

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

Or ComfyUI Manager, search AICG3D / 短剧工作台 / MiniMax H3. Restart, right-click → AICG3D. The pack's only pip dependency is imageio-ffmpeg; H3 itself needs nothing beyond ComfyUI's torch stack.

Weights go where the folder scanners look:

ComfyUI/models/diffusion_models/   # FL2VA / REF2VA weights (unet/ and unet_gguf/ also scanned)
ComfyUI/models/text_encoders/      # qwen3vl_32b_minimax_h3_nvfp4_awq
ComfyUI/models/vae/                # video VAE + audio VAE
ComfyUI/models/loras/

Where people get burned

• One VAE instead of two. The single most common "nothing works" state. You need the video VAE and the audio VAE, and they are different files with different latent dims.
• A GGUF transformer with a mismatched VAE. GGUF weights are in the scanned folders too, so the list will happily offer you a quant of a different model, or the right model with an H3-VAE from the wrong release. When something errors at sample time rather than load time, this is the first thing to re-check.
• The licence, which no error message will mention. H3's community licence carves the EU, UK, Korea and the US out of its applicable territory. The loader will load the weights anyway; it's on you to know whether you're allowed to run them.
• Migrating an old graph. Pre-1.1.x workflows used MiniMaxH3Easy-named nodes. They don't auto-map to AICG3D_H3_*, so a workflow from a tutorial may open with a dead loader node. Replace it by hand.