MiniMax H3-World I2VA 832×480×124 (T8)

This is the node doing the actual magic in the H3-World chain, and it's worth understanding because it's not ordinary I2VA conditioning wearing a costume. In stock MiniMax H3 first-frame I2VA, one prompt conditions the whole clip. Here the node takes your scene prompt, encodes it as the "head," then separately encodes each of the 37 per-latent action sentences from the Action Timeline and binds each one to its own latent with a directed attention mask. Frame 10 can be explicitly walking while frame 30 is explicitly a slow camera pan, because their text never mixes.

The inputs

• clip - the native MiniMax H3 Qwen3-VL CLIP (this is why the node needs that big text encoder in models/text_encoders).
• video_vae and audio_vae - H3 is omni-modal: video and jointly-generated audio. Wire both VAE decoders or the conditioning can't build the av_latent.
• first_frame - your still. Anything that isn't 832×480 gets scaled to cover and center-cropped, per upstream H3-World policy. It's cover-and-crop, never a stretch, so a 16:9 still loses its sides rather than distorting.
• prompt - the scene description. Default is the parking-garage man from the official sample; it's the scene head that every action sentence gets appended to.
• action_plan - the T8_H3_WORLD_ACTION_PLAN out of the Action Timeline node. It's hash-signed, so if you see a validation error here, the plan was modified or it's stale - regenerate it.

The mechanism

Under the hood it encodes the scene head, then walks the 37 sentences from the plan. Distinct sentences are encoded once and cached - "walks forward" appearing on ten latents is one encode, not ten. All the encodings get concatenated into one conditioning tensor with per-segment spans recorded, and the directed-attention metadata tells the sampler which text segment may attend to which latent. That bookkeeping is exactly what MiniMaxH3WorldModelComposerT8Advanced installs the runtime for.

Outputs

The two you wire onward are positive (CONDITIONING) and av_latent (LATENT) - feed both to a native H3 sampler with the model from the Composer. The official workflow runs 50 Euler/native-flow steps with a BasicGuider at CFG 1.0 and video/audio shift 12/3, which is worth copying before you experiment. The other three outputs are for inspection: conditioned_prompt shows the assembled text, action_script_json the 37 sentences, report_json the token/spans diagnostics.

Wiring and expectations

Chain is: Action Timeline → Model Composer → this node → sampler → MiniMaxH3WorldSafeVideoSaveT8Advanced. Grab the sample workflow at examples/workflows/26-h3-world, swap the first frame and prompt, and it should run.

A few things that bite people. Don't change the resolution or frame count downstream - the 37-latent layout and the attention mask are bound to 832×480×124, and the node validates the plan against exactly that. Don't stack OpenVDN/SLA/VSA/Sol-Attn/BlockCache on this chain; this node's conditioning layout is specific and other takeover nodes will misread it. And set your VRAM expectations honestly: H3 is a 33B omni-modal model that's heavy even INT8-quantized, and it wants the text encoder as a second memory budget on top. On 16GB, run one job, and lower the resolution/frame count elsewhere rather than hoping this clip squeezes in beside a second concurrent generation. (Also remember the weight license excludes the US, EU, UK and Korea - verify you're in the applicable territory before downloading anything.)