H3 Continuum Second Pass

If you want to upscale an H3 video but don't want to hand the whole pipeline to the integrated Hi-Res Fix node, H3 Continuum Second Pass V3.5 is the stable alternative. It was built as the Advanced bridge for Issue #8: you upscale the video latent yourself - typically with an external latent processor like LBH - and this node runs one context-aware refinement pass over the result, keeping Continuum's grouping, prompt policy, and your first-pass audio intact.

The README draws a clean line between the two paths. Hi-Res Fix is the one-node Main convenience path (Experimental, and it does its own internal Second Pass). Second Pass is the accepted Advanced path for when the upscale happens outside the pack:

H3 Continuum Sampler V3.5.video_latents
  -> external H3 latent processor (for example LBH)
  -> H3 Continuum Second Pass V3.5
  -> Core Video / Audio VAE Decode
  -> H3 Continuum Assemble + Seam V3.5

How it works

The node reads the physical prompt contract from the assembly plan, then samples each Continuum physical group once at low denoise. It preserves prompt policy, seed reporting, and group order, and it returns the original first-pass audio LATENT objects bit-exact - the refinement touches video only. Like Hi-Res Fix, it has no internal denoise value: the SIGMAS you feed it from a workflow-side BasicScheduler decide how strongly and how many steps the resized latent gets regenerated. The GPU-accepted baseline is res_multistep, simple, 10 steps, denoise 0.35.

Inputs and outputs

Required: model and clip (the same final H3 MODEL/CLIP as the first pass), sampler and sigmas (a separate second-pass set), video_latents (the externally processed ones), audio_latents and assembly_plan from the sampler, and refine_seed (base seed, derived once per physical group). Optional: refine_context (connect the sampler's refine_context so First/Last/Reference conditioning survives at the new geometry) and video_vae.

Outputs: refined_video_latents (list), audio_latents (list, passed through), updated_assembly_plan, and a status string. Wire the latents to Core Video/Audio VAE Decode and the updated plan to Assemble + Seam.

Installing it

Same pack, one install, six nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/ukr8b3g-cmyk/ComfyUI-H3-Continuum.git

Or search "H3 Continuum" in ComfyUI Manager, then restart. No extra pip dependencies.

The gotchas

• Don't stack it with Hi-Res Fix. Hi-Res Fix already performs one Second Pass internally; chaining them double-refines at low denoise and wastes compute for nothing.
• Your SIGMAS are the whole game. There's no denoise slider here. The README's most-important correction exists because people read the baseline as "denoise 10" instead of "10 steps, denoise 0.35." Set the BasicScheduler to simple, 10 steps, denoise 0.35 and refine from there.
• Audio is first-pass only. If you want different audio, that's a job for Driving Audio on the sampler, not this node - the Second Pass deliberately preserves the original audio latents.
• First pass and the 37T groups pass at 1152×1152 on a 16 GB card, but long 2x refinements can still CUDA-OOM at a terminal long group - the README documents exactly that limit. Test short before committing to a long run.