h3-relay

H3 Relay

H3 Relay provides steerable, resumable MiniMax H3 shot generation for ComfyUI. It separates native H3 generation from optional H3 Ultimate or LTX 2.5 enhancement and built-in frame interpolation so creators can approve a shot before paying for finishing work.

Experimental FastH3 / H3 Ultimate support: validated on an RTX 4090 running Linux with the official comfy-kitchen==0.2.33 wheel and the VSA-capable ComfyUI commit pinned in the FastH3 ComfyGit recipe. Stock ComfyUI alone is not sufficient. Windows GPU inference remains unvalidated, and upstream Sol-Attention edge-case tests still have known failures. These limits apply to the experimental FastH3 path, not a new support claim for every H3 workflow. See the validation record.

CI runs unit tests and expanded ComfyUI contracts using CPU-only PyTorch, without model weights. It does not replace the separate GPU generation and audio/video-output validation in the linked record.

The initial node set is:

• H3 Relay · H3 Hybrid Model Loader
• H3 Relay · H3 Model Loader
• H3 Relay · FastH3 VSA Profile
• H3 Relay · LTX Upscale Model Loader
• H3 Relay · Pack LTX Model
• H3 Relay · Cache Manager
• H3 Relay · Apply Model LoRA
• H3 Relay · Attention Backend
• H3 Relay · Sequence Start
• H3 Relay · Generate Shot
• H3 Relay · Accepted Raw Latent
• H3 Relay · H3 Ultimate 2× Enhance
• H3 Relay · LTX 2× Enhance
• H3 Relay · Interpolate
• H3 Relay · Assemble Raw Sequence
• H3 Relay · Assemble

The pack is intentionally staged:

Sequence Start -> Generate Shot 1 -> Generate Shot 2 -> ...
                       |                    |
                       v                    v
             Ultimate/LTX Enhance 1 -> Ultimate/LTX Enhance 2
                       |                    |
                       v                    v
                  Interpolate 1 ------> Interpolate 2 -> Assemble

The Assemble node includes Run staged · bounded RAM. This action derives the dependency order from the visible graph, then queues each H3 shot, enhancement, interpolation, and final assembly as separate jobs. Accepted state is restored from integrity-checked disk manifests, and an explicit release barrier clears ComfyUI's executor cache and loaded models between jobs. A failed stage stops the sequence instead of queueing invalid dependents.

The ordinary ComfyUI queue action remains available for users who explicitly want one monolithic prompt. The staged action is the recommended path for long sequences because its memory requirement is bounded by one active shot rather than growing with shot count.

Raw H3 continuation depends only on accepted raw H3 checkpoints. LTX carries its own 17-frame/three-latent temporal context, and interpolation removes the corresponding duplicated prefix after running. This lets users review and reroll native 480p H3 video/audio before loading LTX or the interpolation model.

The public token names intentionally distinguish the two graph streams:

• sequence is native H3 continuation state;
• enhanced is the Ultimate/LTX/interpolation/assembly state.

Accepted Raw Latent is an optional bridge from sequence to ComfyUI's standard nested H3 LATENT type. It loads an accepted shot's integrity-checked video/audio checkpoint by sequence position (-1 means latest), so latent-aware finishers such as MMH3 Ultimate Upscale can reuse reviewed H3 work without hard-coded cache paths or regeneration. The sequence remains the authority; the node rejects missing shots, missing hashes, and mismatched checkpoint content before returning the latent.

H3 Ultimate 2× Enhance packages that bridge into the same staged finishing contract as LTX. It reuses the connected FastH3 VSA h3_model, automatically reuses the accepted shot's complete global and scene prompt, reconnects up to nine H3 references, expan