H3 LongTake Render

MiniMax H3 gives you a few seconds of video with synced audio, and it's genuinely good - a 33B omni-modal model that treats video and audio as one context instead of bolting a soundtrack onto a silent generator. What it won't give you is a minute. H3 LongTake Render fixes that by slicing your source into overlapping windows, rendering each window as its own H3 clip, hooking each clip to the previous one's latent tail, and writing every chunk to disk the moment it's done. (Caveat: H3's community licence excludes the US, EU, UK and South Korea.)

What it's for

Two jobs, one node. Motion and identity transfer: a reference video drives the performance, and pictures define who's in it (Ref2VA). And style transfer or retexture: re-render the source as Ghibli or GTA, or turn a tank top into gold armour plating while keeping motion and framing - that one wants NRDX's StyleTransfer LoRA. The overlap-into-keyframe trick is what stops the drift. Instead of re-inventing the last frame from a still, the previous clip's latent tail rides across the seam natively.

How it works

The math is easy once you see it. L = clip_frames (always 17k+5 - 124 frames is 5.2s, 243 is 10.1s) and C = context_frames. Clip 0 takes source frames [0, L) and all of them go out; clip i takes [i·(L−C), i·(L−C)+L) and trims its first C frames. You only generate L−C new frames per clip.

Per clip, ffmpeg decodes just that slice - not the whole file - and it becomes <Video 1> (or a guide latent in style mode), while your pictures become <Picture 1..3>. The last C frames of the previous clip, video and audio, are anchored at frame 0 as a native H3 keyframe. Then it samples positive-only (CFG 1 - this is built for the 4-step Turbo LoRA, so don't crank steps), decodes, trims C frames, and writes clip_NNN.mp4 and clip_NNN.latent.pt into output/h3_longtake/<project_name>/.

Between clips the models leave VRAM but stay staged in RAM - that's the performance story, since with the DiT resident the 25 GB Qwen encoder takes ten minutes per clip instead of one. On a 16 GB card at 544×960 with 124-frame clips and 4 Turbo steps: ~4:40 per clip, so a minute of video is roughly twelve clips and an hour.

The inputs that matter

model, clip, vae and audio_vae are plumbing (LoRAs go on model first). source_file is the video in input/; start_seconds / end_seconds trim the range, and ref_image_1..3 are your <Picture N>. Then the handful that decide whether the result is any good:

• source_role - reference (default) treats the slice as a motion suggestion, right for Ref2VA swaps. guide turns it into a guide latent anchored at frame 0: frame-accurate motion and framing, for style transfer with the StyleTransfer LoRA. guide_retexture is that, text-only. guide+reference does both - +50% time for no gain on style transfer.
• context_frames and anchor_mode - 5 and keyframe unless you have a reason. That's the cleanest seam and the best adherence to <Video 1>; context 22 steals attention from the reference, and inpaint at 22 is a visible cut. none is a diagnostic.
• seam_match - turn on color for style transfer. It nudges each clip's first 24 frames toward the previous clip's, dropping the measured seam colour jump from ΔE 5.8 to 0.9.

aspect should stay source - a different canvas makes H3 copy <Video 1> and ignore <Picture 1> - and 0.5 megapixels is about 544×960 in 9:16. Outputs are last_clip, project_dir and report; wire project_dir into the Stitch node.

Installing it

ComfyUI Manager → search H3 LongTake → Install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/mark9009/ComfyUI-H3-LongTake

Restart, and make sure ffmpeg is reachable - the node tries VideoHelperSuite's ffmpeg, then imageio-ffmpeg, then plain ffmpeg on PATH. No requirements.txt, no Python dependencies; all it needs is ComfyUI core ≥ 0.34 plus the models.

The models are the real download: the fp8 Ref2VA diffusion model in models/diffusion_models, a MiniMax H3 Qwen3-VL-32B text encoder in models/text_encoders loaded through CLIPLoader with type minimax (~25 GB, lives in system RAM), both VAEs in models/vae, and the 4-step Turbo LoRA in models/loras.

Where people get burned

Work in the README's order: dry_run = true to read the plan, max_clips = 2 to check identity and the seam, then max_clips = 0, mode = continue for the whole thing. If ComfyUI stops, queue again - it resumes from the first missing clip. One bad clip gets mode = redo_one and a new seed, anchored to both neighbours.

Then the traps. Change the canvas, clip_frames, context_frames, source_role or the source and the node refuses to continue an existing project - use restart or a new project_name. guide without a reference carries no identity. Ref2VA drifts toward <Picture 1> unless the picture matches the video's framing. Never put a style name and <Picture 1> in the same sentence; the model follows the text and ignores the image. And long chains degrade link by link, so restart at a natural cut past a minute.