MiniMax H3 Aicg 已选视频上下文

MiniMax H3's V2V motion transfer was the thing people noticed on launch day - the community's own words were "can do motion control like Kling". AICG3D_H3SelectedVideoContext (MiniMax H3 Aicg 已选视频上下文) is the pack's take on that: take an existing video, slice it into segments on H3's frame grid, and regenerate it with your own references. Replace the person, replace the background, or just re-render the whole thing at a different style and resolution.

It's the entry point of the pack's video-replacement workflow, and it differs from the text-to-video context node in one way that matters: the source video is the timeline master. Segment count, segment lengths and the audio all come from that video, not from a seconds widget.

The one setting that decides everything

video_reference is the fork in the road, and the tooltip is unusually blunt about it:

• On - the original clip goes in as a <Video 1> reference and sampling starts from an empty latent. You can change person and background thoroughly, but the motion is only inferred from the reference, so it may not match the original.
• Off - no reference video; sampling starts from the original clip's own latent. Motion is locked; the person and background get repainted, and how much is decided by the denoise in your schedule. 0.4–0.6 keeps the motion, 0.8 is close to a redraw, 1.0 throws the motion away.

A companion control, source_reference_scale (0–1, default 0.34), only applies when video_reference is on: how far the source is scaled down before being fed as a reference. At 1.0 the model happily copies the original actor's face and clothes. Changed the reference image and the output still looks like the original person? Drop this toward 0.15.

And face_anchor is the identity lock: it crops a face out of your reference image, pastes it onto each segment's source first frame, and feeds that as a keyframe. On, with video_reference on, gives the strongest identity but drags the original silhouette along. Off is the "keep the motion, swap the face" combination - motion anchored by the source latent, identity anchored by the face, denoise 0.45–0.55. (It needs ultralytics and a face_yolov8m.pt in models/ultralytics/bbox/, otherwise it silently falls back to no anchoring and just logs a warning.)

Splitting and shaping

segment_mode is whole_video, time_cuts or frame_cuts, and segment_cuts takes the cut points - the sibling 自动分段切点 node will compute them for you from "every N seconds". context_length (5–141, step 17) is how many frames of the previous segment condition the next one. continuity_mode picks latent_guide / guide / soft_av / hard_av - latent is the lossless cheap one.

There's also a shortcut worth using: replacement_target with replace_people, replace_background or multi_reference generates the prompt for you and switches the plan to per-second slicing with reassembly. replacement_total_seconds limits the replacement to the first N seconds (0 = the whole thing), and replacement_description appends your own constraints to the auto-prompt.

output_resolution and output_aspect_ratio both default to 跟随原片 (follow the source). Pick 720P and 16:9 and you're changing the canvas deliberately instead of inheriting a phone video's 9:16. cleanup_after_run is the finishing move - 卸载模型 unloads and reclaims after 分段解码 finishes, 释放缓存 just hands the cache back (faster next run).

Inputs and outputs

Source media arrives through the pack's resource-library inputs (media, media_1…media_15) - the first video in that library is the timeline master. Required is h3_bundle, then mode (Ref2VA / T2VA / I2VA / FL2VA / L2VA / image / digital_human) and prompt.

Outputs are model and h3_context. The context goes into 分段采样, which auto-detects a selected-video plan and VAE-encodes your source ranges into the same segment intermediate the normal path uses - no denoising happens at this stage. Then 分段二采 if you want a second pass, then 分段解码.

Install

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

Restart and hard-refresh. Face anchoring additionally wants pip install ultralytics plus the YOLO face model; everything else is in the pack. And the usual caveat for anything H3: the local weights are licence-locked out of the US, EU, UK and Korea.

Common issues

Denoise is doing nothing because it's the wrong denoise. With video_reference off, the refine pass prints a hint when your schedule starts above 0.75 that you're about to lose the motion. It's right. Lower it.

"Selected-video segment N has an invalid frame range." Cut points that don't fit the snapped timeline - H3 rounds to a 17k+5 frame grid. Let 自动分段切点 emit the cuts rather than hand-writing them.

Face swap ignored. No detector, no anchor. Check the console line about face_yolov8m.pt.

Beginners tend to reach for video_reference on because "more reference is better". It isn't. If you want the original performance, turn it off and start from the source latent - the whole point of this node is that you get both routes.