H3 Video Outpaint · 生成首帧候选 (EXP)

This is the node I'd point a beginner at first. Outpainting a clip is a long job - the pack's own reference run was 32 seconds of video, 768 frames, 22 serial windows. Sampling all of that and then discovering your geometry was wrong, or that the model decided the new 200px strip is a wall of repeating texture, is an afternoon gone.

MiniMaxH3VideoOutpaintCandidateT8 generates only the first window and shows you the real first frame. Review it. If it's good, connect the selection and continuation nodes and let the rest of the video run under the settings you just approved. If it's not, change something and generate a different candidate. Same configuration, one window's worth of GPU time.

It's an output node with an image preview, which means you can run this node on its own - that's the intended first run, before you wire anything downstream.

What it takes

Required: model, prepared, video_vae, candidate_name, seed, steps, resume, color_match. Optional: geometry_align and source_mode.

candidate_name (default candidate_01) is the identity of this attempt. If you change the seed, the step count or the sampling model, generate a new candidate name - the pack's docs are unambiguous that a changed generation setting needs a new candidate, and the caches are keyed accordingly.

color_match and geometry_align are baked in here rather than chosen at save time, which is the whole point of the design: the settings you reviewed are the settings that get used for the rest of the video. Continuation and saving inherit the candidate's mode and can't quietly switch.

source_mode picks joint_decode (default) or preserve_source. Default joint decode decodes the full canvas together and reconstructs the original region; preserve_source pastes exact source pixels before encoding but can leave a more visible seam. Both keep the existing audio.

resume reuses an already-committed first window if the settings match.

What comes out

Four outputs. candidate is the handle that goes to the confirmation node. first_frame is an IMAGE you can preview or drop into a comparison node. candidate_report is JSON. And candidate_id is a full 64-character identifier - copy all of it, exactly, no truncation, no file paths. That ID is how the loader node finds this candidate again after a restart, and it's the single most-copied string in this workflow family.

How it works under the hood

It runs the first window's sampling and then decodes it through the VAE and composites it with the same geometry and colour policy the final compose will use - so what you're looking at is the actual first frame the delivered video would contain, not a separate preview render. The archive it writes is checked later: the compositing node re-verifies that the recomposed first frame still matches the RGB hash of the preview you approved, and refuses to publish if it doesn't. That's a stronger guarantee than "trust me", and it's why you can't cheaply swap the mode between review and save.

Install

Manager: search MiniMax H3 Audio T8, then restart ComfyUI completely. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

requirements.txt deliberately installs nothing so the pack can't swap out your Torch/CUDA stack. This generation node does need ComfyUI-KJNodes for the audited MiniMaxLowVRAMAttention and MiniMaxChunkFeedForward, ffmpeg on PATH, and the H3 FL2VA model, Qwen3-VL encoder and video/audio VAEs in the documented folders. If the pack's nodes all show red, update ComfyUI, the frontend and Manager, then fully exit and restart - updating the pack alone usually isn't enough.

Two honest limits

The candidate only proves the first window. The pack's released 32-second run passed strict media and audio checks and still had visible strips, repeated textures and seams in some sections. One good first frame is a green light, not a warranty. And don't stack acceleration: this route is Stock20 native noise plus the audited KJ memory patches, and the compatibility audit marks Turbo, SLA, VDN, SPEED and Fast H3 as unsupported rather than merely untested.