3.7 KiB
V3.8 Open Integration Contract
This document defines the supported boundary between H3 Continuum V3.8 and external processing, decoder, and acceleration nodes. It does not add a public node or change the Sampling, State, Session, Run Storage, or Assembly Plan schemas.
Supported pipeline
H3 Continuum Sampler V3.8
-> external latent processor or latent upscaler (optional)
-> H3 Continuum Second Pass
-> Core Video/Audio Decode
-> H3 Continuum Finalize
-> Core Save Video
Pixel upscalers remain outside the latent boundary. If a workflow decodes to pixels for processing, it must encode the result back to one H3 video LATENT per physical group before Second Pass.
External processor contract
An external processor may change only the spatial geometry of the video latents. It must preserve:
- physical-group count and order;
- one indivisible entry for every Long Terminal Merge group;
- video batch, 24 channels, and temporal latent length;
- one common target H/W across every physical group;
- the matching first-pass audio latent list and its order;
- the original Assembly Plan input.
Spatial video dimensions may be preserved or enlarged. Smaller dimensions, NaN/Inf values, changed B/C/T geometry, a changed group count/order, or a changed audio shape are rejected before Second Pass Sampling.
Second Pass rebuilds supported conditioning for the target geometry, samples complete physical groups, and returns a copied Assembly Plan with only target geometry and Second Pass execution metadata added. It does not mutate the input plan or Run Storage. The accepted first-pass audio LATENT objects are returned unchanged; temporary audio produced during video refinement is discarded.
Decoder boundary
H3 Continuum has no decoder allowlist. Any Core or third-party Video/Audio VAE path, including a TensorRT-backed VAE, is compatible when it implements the public ComfyUI contracts and produces the normal decoded types consumed by Finalize:
IMAGE:[frames, height, width, channels]per physical group;AUDIO: a mapping withwaveformand positivesample_rateper physical group.
Finalize depends on those public decoded types and the Assembly Plan, not on a decoder class name. Decoder fallback, tiling, device placement, and acceleration are owned by the decoder or ComfyUI Core.
Acceleration boundary
SageAttention, Sol-Attn, Spectrum, and comparable accelerators remain external
MODEL wrappers. Continuum uses ComfyUI's public ModelPatcher wrapper APIs,
preserves existing wrapper/model options on call-local clones, and publishes
only the documented h3_continuum interoperability hint when continuation
context exists.
Continuum does not select an attention backend, import accelerator-private functions, or globally monkey-patch an accelerator. Missing optional accelerator markers are informational and never become package dependencies.
Ownership summary
| Owner | Responsibility |
|---|---|
| Continuum | Physical-group identity/order, temporal ownership, conditioning reconstruction, Second Pass validation/Sampling, first-pass audio passthrough, target-geometry plan copy, Finalize assembly |
| External processor | Spatial Video LATENT transformation while preserving B/C/T, group order, and finite values |
| Core or decoder node | Video/Audio VAE decode/encode behavior, tiling, fallback, device lifecycle |
| Accelerator wrapper | Sage/Sol/Spectrum backend selection and wrapper lifecycle |
| User workflow | Correctly wire the matching Video LATENT list, first-pass Audio LATENT list, and original Assembly Plan |
The contract deliberately avoids an All-in-One node. New processors and decoders can integrate through public ComfyUI types without requiring a Continuum release, provided these invariants remain true.