AICG-渲染器（二采放大）

What it's for

First pass gets the composition. Second pass gets the detail. AICG-渲染器（二采放大） (AICG3D_H3RenderPass2) takes the pass1 payload from AICG-渲染器（一采） and runs the expensive half: upscale the video latent, re-sample it at moderate denoise so the model fills in real texture, decode, compose. One input, one VIDEO out.

Six canvas nodes' worth of work - split the combined AV latent, upscale the video half in 3D, concatenate back, rebuild the keyframe conditioning at the new resolution, sample, decode - folded into one box. The audio is not upscaled. It rides through unchanged and gets re-muxed at the end, which is the right call: there's nothing to gain from resampling a stereo track through a spatial upscaler.

How it works

The mechanism is exactly the hi-res fix, adapted to a 24-channel 3D latent. The payload's AV latent gets separated into video and audio. The video half goes through a latent upscaler that only changes the spatial grid - the temporal axis is preserved, so your motion doesn't get smeared or doubled. Then it's concatenated back with the audio latent, and the second sampler runs with second_pass_denoise as the strength.

Because a scaled-up latent no longer matches the resolution the keyframe conditioning was encoded at, the pass also rebuilds that conditioning: start/end frame images get re-encoded through the video VAE at the new size before sampling. That's the same job AICG3D_H3SecondPassConditioning does standing alone.

Settings worth your attention

second_pass_denoise - the dial that decides everything. Default 0.5, tooltip says 0.4–0.6 is the stable band. Below 0.4 you've paid for a second pass and got a slightly softer upscale; above 0.6 the model starts rewriting the shot and can eat the detail the upscaler just invented. Start at 0.5.

upscale_scale - 1.2 by default, 1.0 to 4.0. Note that 1.0 means "no upscale, just re-sample at the same size", which is a legitimate way to use this node as a refinement pass.

latent_upscale_model - your checkpoint from models/latent_upscale_models. Empty folder, empty dropdown, and the node errors rather than guessing.

tiled_sampling - off by default for speed. Turn it on for long clips or big upscales: memory is then bounded by one tile rather than the whole frame, at the cost of being a bit slower. tile_width, tile_height (512 each), tile_overlap (128, bigger means softer seams but must stay under the tile dimensions) and tile_fade (32, the actual blend length inside the overlap) tune that. The general rule from every tiled-upscale guide holds: overlap is what stops patchwork.

upscale_device / upscale_precision - cuda / fp16 by default. Drop to cpu or fp32 when the upscaler itself is what's OOMing.

upscale_chunking - on by default, processes long clips in temporal chunks with overlap. Leave it on for anything past a few seconds.

output_first_pass - off by default. Turn it on and you get a decoded copy of the un-refined latent on first_pass_video for A/B comparison. It costs an extra decode, so it's a debugging tool, not a default.

second_pass_seed is 64-bit and 0 means random each run. sampler_name / scheduler / steps describe the second pass, and the optional sampler / sigmas inputs override them the same way they do everywhere else in this pack.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/JGRFW/comfyui-AICG3D

Or Manager → Registry → AICG3D. Restart and Ctrl+F5. The only entry in requirements.txt is imageio-ffmpeg; the latent upscaler is implemented inside the pack, so you don't need a third repo for it.

You do need to supply the upscaler checkpoint:

ComfyUI/models/latent_upscale_models/

The pack creates the folder on demand. Drop the H3-compatible latent upscaler in it and restart so the dropdown picks it up. Everything else - transformers, text encoder, both VAEs - is the same set the first pass uses.

Traps

No upscaler checkpoint. The dropdown shows a single placeholder entry pointing at models/latent_upscale_models, and the run dies with a message telling you to put a checkpoint there. This is the single most common first attempt.

tile_fade larger than tile_overlap. The fade can't be longer than the overlap it lives in; keep fade ≤ overlap and both under the tile size.

Denoise too high. 0.8 feels like it should give "more detail" and instead replaces the upscaled frames with a fresh render that has different faces. If your second pass changed the shot, this is why.

Feeding it a raw latent. pass1 is a payload type, not a LATENT - the node validates it and tells you to connect the first-pass renderer's output. There's no way to use this node as a standalone upscaler.

VRAM. Longer clips are the usual killer, not resolution. Tiled sampling first, then upscale_chunking, then lower the scale.