MiniMax H3 PTD adapters reduce the number of transformer evaluations needed to create video and sound. The setup question is which PTD release you downloaded. EPFL VITA's original four- and six-step checkpoints, a community four- and eight-step Ref2VA implementation, and compact ComfyUI conversions have different base models and loaders.
This guide reviews the official inference release, its ComfyUI instructions, Taobao Content AI's model card, and the compact conversion card, checked October 22, 2026. We have not run these workflows. H3Vid does not currently offer a hosted PTD task.
PTD changes the path through the model
Parametric Trajectory Distillation teaches a student to traverse a generation trajectory in fewer steps. Its research project describes four- and six-evaluation video-and-sound generation. At inference, the adapter used by a layer changes with the hop; it is not equivalent to merging one conventional LoRA and shortening a generic sampler schedule.
The paper's timing comparison excludes VAE decoding and uses eight H100 GPUs. It cannot predict an end-to-end clip time on a single desktop card. The project also distinguishes its selected film clips from a blind preference study. Neither source establishes that every prompt improves over every other accelerator.
Select a distribution before installing
| Distribution | Steps | Base and runtime |
|---|---|---|
| EPFL VITA original | 4 or 6 | Original H3; standalone or dedicated native ComfyUI nodes |
| Taobao Content AI Ref2VA | 4 or 8 | Ref2VA modular Diffusers pipeline and its supplied loader |
| nynxz compact conversion | 4 or 6 | Converted VITA adapters with a compact ComfyUI loader |
The community Ref2VA card demonstrates apply_ptd_lora on transformer_ref. It reads checkpoint metadata to derive the evaluation count, then calls its pipeline with num_inference_steps = nfe + 1. That convention belongs to that implementation. It is not a reason to add an extra step to the official ComfyUI fixed schedule.
Likewise, a community announcement saying “ComfyUI” is not a complete graph contract. Follow the concrete repository for your files. Do not mix a four/eight-step Ref2VA checkpoint with the four/six-step VITA converter or assume their output heads are interchangeable.
Install the official ComfyUI route
The official release documents Python 3.10 or newer and a ComfyUI build with native H3 support. Its qualified baseline is commit 5cc4b870; later builds have not all been qualified. From your ComfyUI directory:
git clone --branch release/comfyui-inference --single-branch \
https://github.com/vita-epfl/PTD-inference.git custom_nodes/ComfyUI-PTD
python -m pip install -e ./custom_nodes/ComfyUI-PTD
Use the interpreter that starts ComfyUI. The validated launch disables DynamicVRAM with --disable-dynamic-vram. Installing only a standalone wheel does not register the custom-node checkout.
The official path needs an original, unpruned H3 transformer converted using scripts/convert_comfy_base.py, plus the separately converted PTD checkpoint in models/ptd/. The guide provides pinned upstream revisions and conversion commands. Keep the DCP checkpoint metadata and all shards together during conversion; the runtime uses the resulting safetensors file.
Add the native H3 text encoder, video VAE and audio VAE in their documented directories. The example filenames for converted base and adapter outputs are local choices, whereas the encoder and VAE names identify actual published files. Do not search for an invented upstream download called minimax_h3_original.safetensors when the guide tells you to create it.
Use the PTD graph, including its sound path
Load examples/comfyui/ptd_h3_t2v.json. The core route is the diffusion loader, PTD adapter loader, Basic Guider, PTD Sampler and PTD Sigmas into SamplerCustomAdvanced. Split the resulting AV latent, decode its video and audio with their matching VAEs, and combine at 24 fps before saving.
Connect the same adapted model to the guider, sampler and sigma nodes. Basic Guider supplies no additional classifier-free guidance. Keep an empty generated latent and the complete checkpoint schedule. Generic schedulers, truncated schedules, denoise masks and prefilled generated latents are outside this release's documented support.
The official full-model validation covers four-step text-to-video, including encoding, sampling, decoding and saving. Six-step and image-conditioned ComfyUI paths are implemented but remain unvalidated on the full model. A backend queue run is also different from manual browser testing or a listening evaluation.
Compact files have a separate compatibility boundary
The nynxz conversion offers BF16 and smaller INT8 files. Its compact loader is required for the INT8 files, pruned bases and its DynamicVRAM support. It describes FL2VA as the matching original base and rejects Ref2VA, while its table labels several quantized bases as loading but untested.
Its measured adapter error is a weight-level calculation, not a video-quality benchmark. The maintainer explicitly says no end-to-end generation was run with those compact files. Treat smaller downloads as an engineering option with their stated evidence boundary, rather than guaranteed equivalent output.
Start with the official four-step graph if reproducibility is the priority. Move to compact files only after recording the loader, base revision and intended memory tradeoff. For a different acceleration family, consult FastH3 Omni setup; it is a separate checkpoint and execution recipe.
The cover is an imagegen concept illustration, not a PTD output, measured trajectory or software screenshot.