MiniMax H3 在 ComfyUI 的十套官方工作流涵蓋四種影格路線及六種 R2V 變體

MiniMax H3 is one model that covers a lot of ground, and this episode's ten graphs are the cleanest way to see how much of that ground is just different conditioning. Four frame routes (text, first frame, last frame, first-and-last) and six reference-to-video variants (multi-image, pure video, character + motion video, character + standalone audio, character + video with paired audio, character + motion video with new audio).

No inference was run for this release. Everything below is read from the saved graphs. There is no timing, VRAM or quality claim.

One attribution note first: every graph's Markdown note links the official Comfy-Org/MiniMax-H3 model files and the MiniMax blog. Those notes are preserved in the distributed JSONs; this article is documentation of the templates, not a claim to have authored them.

The shared spine

All ten load minimax_h3_fl2va_bf16.safetensors, encode text with qwen3vl_32b_minimax_h3_int8_convrot.safetensors (type minimax) and decode through minimax_h3_video_vae_fp16.safetensors plus minimax_h3_audio_vae_fp32.safetensors. There are no LoRA nodes anywhere in the set — this is the base H3 stack, not a distilled variant.

Sampling is identical in all ten: KSamplerSelect res_multistep with BasicScheduler simple at 20 steps and denoise 1, executed through SamplerCustomAdvanced. Output always goes to the SaveVideo prefix video/MiniMax_H3.

What changes is the conditioning node and what you feed it. MiniMaxH3ImageToVideo handles the frame routes; MiniMaxH3ReferenceToVideo handles the R2V routes and is saved at 1344x768 / 124 / match.

The frame routes: 02 is the baseline

Start with 02 (text-to-video). It has no image input at all — 23 nodes, one branch, a 10-second 16:9 prompt. Everything else in the frame group is that graph plus an anchor:

• 04 adds one image as the starting frame.
• 01 uses one image as the ending frame, and it is the odd one out structurally: two parallel branches (10 s and 5 s) live in the same file, each with its own loader stack. That is convenient but it doubles the node count, so read which branch is wired to which save before you edit.
• 03 takes both a start and an end image and asks the prompt to preserve the same subject across the gap.

If your result drifts in the middle, the fix is usually in the prompt's "preserve" clause, not the sampler. These are 20-step base-model graphs; there is no turbo LoRA to retune.

The reference routes: read the input list, not the title

The six R2V files look alike in the node list but differ in which reference material is wired in:

• 05 is the simplest: four still images plus one audio clip (with a trim node). Good for character boards and product angles.
• 06 is the video-reference route: LoadVideo → Video Slice → GetVideoComponents. The slice node is what decides which segment of the reference clip is used.
• 07 combines both: character stills for identity, a motion video for movement.
• 08 drops the video and keeps a standalone audio clip — this is the local-only graph in the set, with no RunningHub page in the source material.
• 09 and 10 are a pair worth understanding together. 09 uses the reference video and its paired audio; 10 replaces that audio with a new standalone clip. If you are doing a dub or a voice replacement, 10 is the one you want.

Practical notes before queueing

• Check the duration widget per file — the set ranges from 4 s to 10 s and the saved values are not uniform.
• The R2V routes have an audio trim node; if the output cuts speech, check the trim range before changing anything else.
• 01 has two branches in one file. If you edit the prompt and only one output changes, you edited one branch.
• Frame routes at 0.8 MP and reference routes at 0.4 MP are different working sizes; do not compare outputs across the two groups as if they were the same experiment.

What this is not

Not a benchmark, not a quality review, and not a license grant. The video's framing about "one model doing everything" is the author's description. The graphs' notes credit the official Comfy-Org/MiniMax-H3 model files; the files themselves grant no rights to the H3 weights, the encoder, the VAEs or the source media. Voice and likeness use have their own legal requirements.

Getting the workflows

Published by this channel. The graphs are free; the H3 weights, encoder, VAEs and example media are not bundled.

• Hugging Face: AI-KSK/minimax-h3-ten-official-workflows-288 — ten JSONs, source covers, SHA256 manifest. Start here.
• ModelScope (魔搭): Sunjian520/minimax-h3-ten-official-workflows-288
• Civitai: 01 last frame · 02 text-to-video · 03 first+last · 04 first frame · 05 R2V-01 · 06 R2V-02 · 07 R2V-03 · 08 R2V-04 · 09 R2V-05 · 10 R2V-06
• Run without a local GPU: 01 · 02 · 03 · 04 · 05 · 06 · 07 · 09 · 10 — these carry the author's invite code rh-v1111. (08 is local-only.)
• Tutorial: YouTube · Bilibili

Verify workflow, node, model, encoder, VAE, source-video/audio and cover rights separately. No blanket commercial or redistribution license is asserted for the graphs or the example media.