MiniMax H3 Video Extend — Full Set — August 2026 [For Patreon Supporters] (Patreon)

Hey everyone.

If you watched the long-link video extension tutorial, you saw the two-segment forest demo and the basic extend concept. (Public Post: https://www.patreon.com/aifuturetech/posts/minimax-h3-long-168038037)This post has the production set — the four-segment extend chain, and the versions with the H3 latent upscale stage, so you go from 0.8MP segments straight to high-res output in one graph.

----------------------------------------

WHAT'S INCLUDED

----------------------------------------

1. H3 Turbo R2V + Extend + Latent Upscale (2 segments)

.....................................

File: MiniMax-H3-Turbo-R2V-Extend-LatenUpres-202608.json

The workflow from the video's high-res demo, end to end. Segment one generates with native MiniMaxH3ReferenceToVideo (turbo 4-step LoRA at 0.7). Segment two runs MiniMaxH3VideoExtendPatched — it takes the context latent from the first video plus your two character reference images and the audio reference, and continues the motion. After stitching, the MMH3UltimateUpscale stage (my resident-model fork) re-samples the latent up with minimax_h3_latent_upscaler_3d_bf16, with spatial/temporal split params (H3_SPATIAL_PARAM / H3_TEMPORAL_PARAM / H3_UPSCALE_PARAM) exposed so you can tune chunk sizes for your VRAM.

Use it when you just want one clip at high res, or as the "segment 1 factory" when you run the extend stages as separate workflows (the API / agentic split I show in the video).

2. H3 Turbo R2V + Latent Upscale (generation only, no extend)

.....................................

File: MiniMax-H3-Turbo-R2V-LatenUpres-202608.json

Same graph minus the extend node — this is the pure first-segment generator plus the upscale stage.

Use it when you just want one clip at high res, or as the "segment 1 factory" when you run the extend stages as separate workflows (the API / agentic split I show in the video).

3. Ref2V Video x4 Extend (4-segment chain, base res)

.....................................

File: Minimaxh3-Ref2V-video_4extend.json

The four-segment chain from the second half of the video. Each segment's latent sampler output (latent two, latent three...) feeds the input latent of the next MiniMaxH3VideoExtendPatched block, with its own text prompt per segment. Reference images (character sheets) and audio reference carry through all four. Nine VHS outputs so you can inspect every segment and the stitched result. No upscale — fast iteration at 480p / 0.8MP.

4. Ref2V Video x4 Extend + Latent Upscale (the full pipeline)

.....................................

File: Minimaxh3-Ref2V-video_4extend-LatentUpscale.json

Everything above plus the latent upscale stage on the final stitched output. This is the one to run when you're done testing motion and want the keeper.

----------------------------------------

WHAT YOU NEED

----------------------------------------

Model files (exact paths):

models/diffusion_models/minimax-h3/minimax_h3_ref2va_int8_convrot.safetensors

models/loras/minimax-h3/minimax_h3_ref2v_turbo_4step_v0.1_comfyui_bf16.safetensors

models/text_encoders/minimax-h3/qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors

models/vae/minimax-h3/minimax_h3_video_vae_fp16.safetensors

models/vae/minimax-h3/minimax_h3_audio_vae_fp32.safetensors

models/upscale_models/minimax-h3/minimax_h3_latent_upscaler_3d_bf16.safetensors (upscale versions)

Model sources:

- All H3 models: https://huggingface.co/Comfy-Org/MiniMax-H3/ (direct resolve links are also pasted inside the workflow notes)

- Text encoder alternative in the graphs: qwen3vl_32b_h3_ultra_uncensored_heretic_int8_convrot.safetensors https://huggingface.co/ethanfel/Qwen3-VL-32B-Ultra-Heretic-H3-ComfyUI-INT8-ConvRot

Custom nodes:

- ComfyUI-MiniMax-H3-Extend — Kayte's node (MiniMaxH3VideoExtendPatched) https://github.com/kat3ri/ComfyUI-MiniMax-H3-Extend /

- ComfyUI-MiniMaxH3 (native H3 nodes: ReferenceToVideo, SigmaShift, encoders)

- MMH3UltimateUpscale + split param nodes (my fork from the Latent Upscale video — same node pack)

- ComfyUI-KJNodes (ImageResizeKJv2)

- ComfyUI-VideoHelperSuite (VHS_VideoCombine)

- ComfyUI-Custom-Scripts or equivalent for GetNode/SetNode + ComfyMathExpression

GPU:

- 24GB+ VRAM recommended for the extend + upscale graphs. The 4-segment base-res chain runs on less; the upscale stage is what eats headroom.

----------------------------------------

QUICK START

----------------------------------------

1. Install the custom node packs, drop the models in the paths above.

2. Start with file 3 (4extend) at 480p. Load your two character sheets as the reference images in every segment block.

3. Write prompt 1 as the opening action, and each next prompt as a CONTINUATION ("...the two characters keep running and start fighting zombies"). Each segment inherits the previous latent, so don't re-describe the whole scene — describe what happens NEXT.

4. Keep each segment at 11-12 seconds.

5. Wire check before queueing: segment N's input latent must point at segment N-1's sampler output (latent two → third block, latent three → fourth block).

6. When the motion reads right, move to file 4 and set resolution by megapixel: 0.8 MP = 1216x672 (16:9).

7. Trim the first second or so of each extended segment in your editor — it removes the overlap/stitching frames and the result plays smoother.

----------------------------------------

TIPS

----------------------------------------

1. Turbo LoRA at 0.7, not 1.0. Same rule as every other H3 workflow — full strength makes motion aggressive and hurts consistency across segments.

2. The audio reference is not optional if you care about seams. It's what keeps segment two from sounding like a totally different recording. Feed the same audio reference through all segments.

3. Chain latents, don't save latents to disk. Some setups index latent files on disk and re-read them — naming conventions drift and it breaks. Wiring the output latent node directly into the next extend block is more flexible and it's what lets you move the whole thing to an API/tool call later.

4. Multi-camera shots still work inside extended segments — that's the real test. If your environment holds from segment one's background all the way to segment four with different angles, the chain is healthy. If it drifts, your prompt is re-describing the scene instead of continuing it.

5. Steady single-camera shots (one person running/dancing) are easy mode — this pipeline is really for multi-shot sequences. Start there only if you're validating the wiring.

6. Duplicate a segment block to add segment five, six, etc. The pattern is mechanical: new extend node, new prompt, point the input latent at the previous block's output.

----------------------------------------

COMMON ISSUES

----------------------------------------

Second segment's motion jumps at the seam

→ You're feeding the wrong latent. Check the extend block's input latent matches the PREVIOUS segment's sampler output, not the first segment's.

Backgrounds change between segments

→ Your continuation prompt is re-describing the environment. Only describe the new action and camera move; the latent carries the scene.

Audio feels chopped between segments

→ The audio reference isn't connected through the extend block, or it's a different clip than segment one's.

VRAM overflow on the upscale stage

→ Lower the spatial/temporal split chunk sizes in the H3 param nodes, or run the 4extend (no-upscale) file first and upscale as a separate pass with file 2.

Extend node missing / red after import

→ Update both node packs. MiniMaxH3VideoExtendPatched comes from the H3-Extend repo (Kayte's or my fork), not from ComfyUI core.

Overlapping frames visible at the join

→ Expected — trim the first second of the extended segment in post, like I do in the video examples.

----------------------------------------

The workflow files are attached. Install the custom nodes first, then check the model paths. Everything else is in the graphs.

Huge thanks to Kayte for building the extend node — she made this whole video possible.

Questions? Drop them in the comments and I will help you sort it out.

Thanks for supporting the channel. You guys make these deep dives possible.

— Benji