MiniMax H3 long videos in ComfyUI: retake one chunk, keep the rest (Ultimate H3 0.13)

Stubelius Ultimate H3 0.13, my free MiniMax H3 workflow for ComfyUI, renders a long video chunk by chunk and keeps every chunk as a take, so you can redo one part alone. On an RTX 5090, retaking chunk 2 of a 12-second Speed video at 864×480 took 53.7 seconds, against 83.2 to 120.9 for rendering both chunks.

Stubelius Ultimate H3 is my MiniMax H3 workflow and custom node pack for ComfyUI. In 0.13 it is four windows: Setup decides how the video is made, the Director what it is, Seeds plays the candidate videos side by side, and Output finishes the one you pick.

How does Ultimate H3 make a long MiniMax H3 video?

It renders the video in chunks, one MiniMax H3 generation each, and every chunk after the first starts from the one before it, unless it cuts to a new shot. In the Director you set the length, up to 120 seconds, and the chunk size, 3 to 15 seconds (10 by default).

For the smoothest joins the pack recommends seitanism's ComfyUI-H3-Motion-Context-MultiRef. Each chunk also has its own shots, style, soundscape, music and LoRAs, while the turbo LoRA and two global LoRAs in Setup apply to every chunk.

How do you retake one chunk without rendering the whole video again?

Pick the chunk in the Director and press Retake. It gets a seed of its own and the video renders up to it, while the chunks before it load from their takes. On my RTX 5090, retaking chunk 2 of a 12-second Speed video took 53.7 seconds, against 83.2 to 120.9 for both chunks.

A take is found by what made it, from the prompt and settings to the seed and the chunk before, so a chunk that would come out the same is loaded, not sampled. A loaded take matched the render within one 8-bit step on every frame but the two blended at a join, with bit-identical sound.

Tested 2026-10-15 on an RTX 5090 (32 GB), 96 GB RAM, Windows 11. Speed mode (8 steps, 0.4 megapixels: 864×480 at 16:9, turbo LoRA at 1.0), First/Last Frame mode with no pictures, one text-only scene, seed 7 (8 in the last row). Render only, no finish.

• Takes, under Retake, lists the chunk's kept takes: click one to pick it back.
• Stop keeps the chunks it finished. Chunks after a retaken chunk that start from it render again with the next Render.
• A different seed at a join can change the fine grain. If a retake shows that, retake again or pick the take before.

How do you hunt seeds for one chunk instead of the whole video?

Press Hunt on that chunk, with the arrow beside it set to 2, 3 or 4. The candidates differ in that chunk alone, the chunks before it render once, and you choose one in the Seeds window with Pick by chunk. Its seed becomes the chunk's own, and you can hunt the next chunk on top of it.

I built it because I often wasn't happy with how chunk 1 ran into chunk 2, and starting the whole video again got frustrating. Seeds to compare in Setup still renders 1 to 4 full videos, and Render the rest renders the chunks after your picks.

The Seeds window plays every seed side by side and in sync. Solo one seed's sound, or you hear them mixed. Then Finish the winner. End to end, two seeds of the 12-second Speed video took 3 min 13 s, and the finish to 1080p at 30 fps with RTX VSR another 15 s.

How many middle frames can MiniMax H3 hit between the first and last frame?

In Ultimate H3's First/Last Frame mode, up to eight, plus the first and the last frame, in one shot. Each is pinned on its own frame with ComfyUI's Add Guide for MiniMax H3. In my tests for 0.12 (2026-10-08), eight of eight landed on their exact frames.

Middle frames sit on the Keyframes lane, at least half a second apart. The text encoder sees only the first and the last frame, so write the way to each middle frame in the shot. On a 0.12 test build that took more (2026-10-08), 1 to 38 middle frames in one 6.5-second chunk all landed on their own frames, with no jumps or holds. Don't pin fast action in the first half second of a chunk that continues the video: it can snap into place for a frame.

Tested 2026-10-08 on an RTX 5090 (32 GB), for 0.12: PDMD mode at 1344×768, one 6.5-second chunk, pictures cut from a render of the same scene and pinned in a render with another seed.

In the video's showcase shot, one 10-second PDMD render through a first frame, a middle frame at 4.5 seconds and a last frame, the middle frame landed on its exact frame.

The Sound lane pins up to 8 sounds or short clips at their seconds. A sound doesn't move the picture: for a reaction on it, pin a middle frame just after, as in the 0.11 door-slam test.

How do you keep a character's face the same across chunks?

Make a character from 8 to 16 pictures on the Characters lane and keep it on for the whole video. In First/Last Frame mode, a character from 8 pictures scored a face match of 0.74 against 0.40 for words alone, as high as the character's own footage, in a 6-second Hybrid test on three seeds.

Use the face from different angles and light, and describe the character in words too: H3's text side never sees the pictures. The files are the RefMods of Luisacaotica's ComfyUI-MiniMaxH3Mod, so they load in both packs.

Tested 2026-10-15 on an RTX 5090 (32 GB). Hybrid mode, one 6-second chunk per take, seeds 101, 202 and 303. Face match is face recognition against other footage of the character.

It held over a join too: in 12 seconds in two chunks, every seed's second chunk scored 0.77 to 0.82 with the character and 0.33 to 0.62 without. Eight pictures at 0.5 megapixels each (the default) added about 4 seconds to a 6-second Hybrid chunk; at 1 megapixel, 13 to 16. In Reference (Omni) mode, use a cast picture and a character together.

Which mode should you pick in the Setup window?

PDMD for most shots: 4 steps at about 768p, close to the 20-step quality, with caveats in motion and fine detail. Speed and Hybrid suit quick drafts with the turbo LoRA, and Quality, 20 steps with a polish, suits shots that must hold up. Setup also holds the seed count, the models, the speed-up LoRAs and the low-VRAM switches.

PDMD is measured in the PDMD test; its LoRA comes from pdmd2026/pdmd_4NFE_lora, as published.

What does the Output window do?

It finishes only the winner: final size, a frame-rate slider from 24 to 120 fps (RIFE draws the frames in between), RTX VSR or DLSS 5 with Color Lock, de-stutter and the Quality polish. Changing anything there re-runs only the finish, never the seeds.

In the video I finish a 10-second Hybrid shot to 1080p at 60 fps with DLSS5 + Color Lock, which keeps the original colours (DLSS 5 against RTX VSR). A split view puts the winner against the finished video, with a loupe for detail, and a Last frame card saves the final frame to start the next video from.

De-stutter, which I made with anime and 2.5D in mind, fixes takes that judder, where the picture holds a frame in a steady rhythm: RIFE draws the held frames again, keeping the length and the sound. In my 0.11 test (2026-10-07), on a 15-second Speed take it found all 95 held frames, and the motion's roughness fell from 0.83 to 0.23. Leave it on auto, which acts only on takes that judder.

How do you update to Ultimate H3 0.13?

Pull the latest main from GitHub or update through Manager, restart ComfyUI, and open the new example workflow to get the four windows. There is no release zip: 0.13 is on the repo's main branch, and your older Ultimate H3 workflows still open.

1. In ComfyUI/custom_nodes/Stubelius-Ultimate-H3, run git checkout main and git pull, or Manager → Update All.
2. Restart ComfyUI and reload the page.
3. Open the example workflow under Workflow → Browse Templates → Stubelius-Ultimate-H3 to get the four windows.

For a new install, clone the repo into custom_nodes or use Manager → Install via Git URL. It needs ComfyUI 0.34 or newer. New to ComfyUI? The free Part 1 of ComfyUI From Zero covers the install.

Test setup and limits

• One RTX 5090 (32 GB), 96 GB RAM, Windows 11.
• The takes timings are one run each of one 12-second Speed video, not PDMD or Quality.
• The face matches are one character in one scene, three seeds per row.
• The middle-frame numbers come from 0.12 tests (2026-10-08), the de-stutter numbers from one Speed take in 0.11 (2026-10-07), and the showcase shot is one render.
• The 2026-10-15 tests ran on 0.13 test builds, with a community variant of the text encoder in the same int8 format.
• I didn't benchmark the video's render or test cards under 32 GB.

Sources and files

• MiniMax H3 Long Videos in ComfyUI: Retake Any Chunk, Keep the Rest (21:59): the modes from 1:46, hunt one chunk from 7:51, middle frames from 9:34, characters from 10:19, takes and retakes from 12:43 and the Output from 15:29.
• stuubszzz/Stubelius-Ultimate-H3: the pack, example workflow and README (MIT), built on Muse Minimax Director V1.2 by Muse Collective.
• Luisacaotica's ComfyUI-MiniMaxH3Mod (the RefMods format), Fannovel16's ComfyUI-Frame-Interpolation (RIFE) and Blueforcer's ComfyUI-DLSS5-Enhancer.
• The MiniMax H3 model card. The model is MiniMax's, under the MiniMax H3 Community License.
• On this site: the first Ultimate H3 guide, written for 0.8, and middle frames, clips and pinned sounds.

Common questions