CineStack H3 Director Refine

H3 at 864×480 is a preview canvas, not a deliverable. You have two ways to get from there to something you'd show someone: run a second sampling pass at the same size to clean up what the first pass fumbled, or enlarge the canvas and sample into it. CineStack H3 Director Refine does both, plus a third thing that isn't sampling at all.

It plugs into the Director's refine socket. Unconnected, the Director is a single-pass sampler and nothing changes. Connected, every segment you run gets a second pass, and the Director's images becomes the refined result while images_pre_refine preserves exactly what the first pass produced - so you can actually A/B the thing instead of guessing whether the second pass helped.

The three modes

mode: refine is the same-resolution second sample. Think of it as giving the model another crack at the same composition with a fresh sigma schedule; it's the mode that fixes mushy motion and soft detail without changing your shot size.

mode: upscale enlarges to a target canvas first, then samples. This is hi-res-fix logic from the image world carried into video, and it's the mode most people want for delivery. upscale_method decides how it enlarges: h3_latent does it in H3's own 3D latent space with a matching 3D weight, lanczos is pixel interpolation (optionally through upscale_model, e.g. RealESRGAN_x2plus), and nvidia_rtx_vsr hands it to NVIDIA's RTX Video Super Resolution, which needs nvidia-vfx and an RTX card. The pack deliberately does not install that for you.

mode: latent_upscale enlarges the latent and doesn't sample afterwards. Fast, cheap, and honest about it - you get a bigger frame without new detail invented for it.

Sizing and sampling

megapixels is the dial for enlargement, not aspect_ratio - that field is deprecated and the second pass always follows the Director's first-pass aspect. A 864×480 first pass (16:9) at 1.8MP lands at 1824×1024, per the tooltip's own example. Leave aspect_ratio alone; width/height only matter in the custom case.

passes is how many refine iterations run; with upscale, only the first one enlarges and the rest are same-size refinement. latent_upscale ignores it.

For sampling, sampler defaults to euler, the example template's choice, while the author's note says high-quality second passes commonly pair res_multistep with a BasicScheduler. Which brings up the rule that trips everyone: sigmas must be wired for refine and upscale. Feed it from Comfy's BasicScheduler or ManualSigmas, and connect that scheduler to the same MODEL the second pass uses - the Director's model, or refine_model if you plugged a different UNET in. H3's SigmaShift is applied inside Refine, so don't stack another one on.

seed_mode resolves to inherit (follow the segment's first-pass seed), offset (that seed + 1), or independent, which uses the node's own seed with the usual fixed/increment/decrement/randomize behaviour and never touches the Director's seeds.

Then the memory knobs, which matter more than they look: enable_tiling with tile_count and tile_overlap splits the spatial pass; enable_latent_chunking slices long latents into 24-frame chunks with weighted overlap so a long segment doesn't die in the upscale step. Both trades are the familiar one - lower peak VRAM, more forward passes, and a seam that may not match a whole-segment forward pass exactly.

Outputs, plus the two switches that bite

refine goes to the Director. width and height are the resolved target canvas (multiples of 32) and are handy for a display node or for sizing a downstream post chain to match.

skip_fl2v defaults to true, and that's the right default: a second pass repaints, and on a first/last-frame shot the pinned keyframes are the whole point. Turn it off only when you've decided you want new pixels at the ends.

confirm_first_pass is the "look before you leap" switch. On, the first Queue with no matching first-pass cache runs only the first pass and writes the cache plus a _pre.mp4; the next Queue with the same seed skips straight to refinement. It only skips when the fingerprint genuinely matches - same seed, same first-pass settings, same SelfLift and Semantic Bridge wiring and values. Change anything and it won't skip, which is the behaviour you want even though it feels like it isn't.

refine_only is the strict version: check every selected segment's first-pass cache, list the mismatches by segment and field, and stop. It never silently re-runs a first pass. Only valid for refine and upscale.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/guomenghao/ComfyUI-CineStack-H3.git
python -m pip install -r ComfyUI-CineStack-H3/requirements.txt

For h3_latent and latent_upscale you also need a 3D upscale weight in ComfyUI/models/latent_upscale_models/ - the dropdown lists whatever H3-family files it finds there. nvidia-vfx is not in requirements (AMD and cloud installs would break); install it yourself for the VSR path.

Where people get burned

Nothing happens, no second pass. sigmas is unconnected. In refine and upscale it's mandatory - steps, scheduler and denoise widgets are all ignored when it's absent.

"Second pass only" refuses to run. By design. Some segment's cache doesn't match its current prompt, seed or wiring, and the pack tells you which. Fix the mismatch or re-run the first pass for that segment.

A visible band or shimmer after enabling chunking or tiling. Overlap-weighted chunking is not identical to a whole-segment forward pass. If you're chasing the last few percent, turn the chunking off and pay the VRAM.

You upscaled with latent_upscale and it's still soft. Enlarging a latent adds no information. Want detail? You need the sample - upscale with h3_latent, or a real video upscaler in post. Keep the second-pass original and post-process from a copy, so you can tell which stage changed the picture.