video_minimax_h3_t2v.json explained

This guide is for ComfyUI users who want the exact prompt-only graph, its required files and a measured baseline before they commit to a full local T2V run.

The pinned graph leaves both first_frame and last_frame disconnected. Your prompt supplies the shots, motion and sound direction; the H3 node produces a video stream and a native audio track, which the SaveVideo node writes together.

That makes this the clean starting point when there is no source image whose composition must survive. If one image should anchor the first frame, use the I2V template. If reference media should guide identity, style or motion, use the R2V template.

Why freeze it? The upstream file can change under the same name. Pinning the revision makes the graph you download the same graph whose loaders, defaults and SHA-256 this page describes.

These are evidence plates, not decorative mockups. The first maps the three live links in the committed JSON; the second copies the two retained run records and their shared output hash from the public ledger. Neither image contains MiniMax-generated output.

Evidence sources: the committed UI graph, public run ledger and the complete FL2VA baseline record.

The local T2V and I2V templates both load minimax_h3_fl2va_pruned_int8_convrot.safetensors. The R2V template does not: it loads a separate Ref2VA file. If you already installed I2V, T2V adds no second diffusion-model download.

The full six-template map and the two-step disk plan live on the workflows hub. This page stays with the one graph it measured.

ComfyUI/
└── models/
    ├── diffusion_models/
    │   └── minimax_h3_fl2va_pruned_int8_convrot.safetensors
    ├── text_encoders/
    │   └── qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors
    └── vae/
        ├── minimax_h3_video_vae_fp16.safetensors
        └── minimax_h3_audio_vae_fp32.safetensors

The pinned template names all four files directly. Before the benchmark, the same file names were checked against the test machine's model manifest; no Turbo LoRA, cache node, SageAttention package or custom node was present.

The Qwen3-VL encoder's upstream project is Apache 2.0 — read the upstream Qwen3-VL license — while the Comfy-Org repackage is delivered through a repository marked with the MiniMax H3 community-license metadata. Download entrance matters; the file-to-license map keeps those tracks separate.

Template and loader source: the pinned Comfy-Org JSON, checked 2026-08-21. Folder placement and loader troubleshooting are covered once in the ComfyUI setup guide.

Full record: download the complete FL2VA T2V/I2V baseline report, including asset fingerprints, workflow hashes, smoke runs, both formal runs and retained evidence locations.

These are two observations from one fixed machine, not a specification for every RTX 3060. Both original timings remain visible; neither is replaced by a single summary value.

With no source image to resize, the test held the prompt, seed, model, sampler and audio path fixed and changed the output workload. The smoke test used 512×288 and 22 frames; the formal pair used 1344×768 and 124 frames.

The formal workload contains about 39.5× as many pixel-frames as the smoke test and took about 42.0× as long. Peak VRAM did not move in the same direction: all three readings stayed between 10,863 and 11,625 MiB. On this workload, reducing the canvas bought time, not proportional graphics-memory headroom.

Duration is the second workload control, but it does not translate into an arbitrary frame count. The pinned graph converts seconds to H3's valid 17k+5 frame grid at 24 fps; a five-second entry becomes 124 frames. Shortening the value therefore moves between valid blocks rather than subtracting frames one by one.

The file also exposes one positive prompt and no negative-prompt socket. For a clean first diagnosis, leave that prompt and the seed fixed while changing only ResolutionSelector or duration. A source image cannot be the hidden variable on this graph, because neither frame input is connected.

The selector bundled with the pinned JSON starts at 16:9, 0.4 megapixels and a multiple of 32; its own reference table maps that preset to 864×480. The 1344×768 figures above came from a separately frozen benchmark setting and must not be read as the template default. For the memory mechanism, see the VRAM guide; for the 43GB-class host-memory peak, see the system RAM guide.

Default-setting source: the pinned Comfy-Org template and its bundled size table, checked 2026-08-21. The smoke and formal values are site measurements recorded under separate test_id values in the run ledger.

This T2V pair is the unaccelerated denominator for the site's same-machine Turbo LoRA test. The dedicated /loras/turbo guide will own the LoRA file names, step counts, shift settings and A/B conclusion; until that page ships, they are deliberately not duplicated here.

The underlying Turbo records already live in the benchmark ledger. Keeping the interpretation off this filename page prevents a baseline download guide from turning into a second accelerator guide.

The measured 36-minute pair is a planning reference for this one RTX 3060 setup, not a promise for another card or host. If a first image must anchor composition, stop here and switch to the I2V graph; if reference media should guide identity or style, use R2V.

Your next action is concrete: download the pinned T2V JSON, verify its SHA-256 and run the small test before changing prompt, duration or canvas.