FastH3-ComfyUI-Community-Workflow

FastH3 ComfyUI Community Workflow

An experimental MiniMax H3 workflow based on a locally tested production graph. It combines a low-resolution FastH3 first pass, learned latent upscaling, an optional Ref2VA refinement pass, NVIDIA RTX video upscaling, and one-prompt-per-image multi-folder routing.

The v0.1.0 workflow ZIP contains no model weights. The repository also provides fasth3_dense_datafree_6step_pruned_comfyui.safetensors as a separate download. This FastH3 LoRA was not trained by this project: it is an unofficial ComfyUI conversion of FastVideo's public dense-datafree adapter, made with NikoDemon80's public converter. Keep the upstream credits and follow the MiniMax H3 license and all applicable upstream terms.

Territory notice: the current MiniMax H3 Community License excludes the European Union, United Kingdom, Republic of Korea, and United States, and restricts use, running, distribution, or display of MiniMax H3 Works and their outputs in those territories. Read the included MINIMAX_H3_LICENSE.txt before downloading the LoRA or publishing outputs. The v0.1.0 workflow ZIP contains only independent workflow files, prompt text, and original node code; the converted LoRA is a separate repository file. This is not legal advice.

Included

• workflow/FastH3_4step_10s_batch_manual_prompt.json: public, portable workflow using a core multiline string node.
• custom_nodes/ComfyUI-H3-MultiFolder-Router: batch input/output routing nodes.
• prompts/childwear_10s_4shot_director_cn.txt: a Chinese director instruction for a vision-language model to turn an input image into a finished H3 prompt. It is not itself a ready-to-run H3 prompt.
• MODELS_AND_DEPENDENCIES_CN.md: exact files, locations, and upstream links.
• NOTICE.md: provenance and licensing notes.
• LORA_PROVENANCE.md: exact LoRA source, conversion changes, checksum, authorization statement, and installation settings.
• MINIMAX_H3_LICENSE.txt: a copy of the MiniMax H3 Community License retrieved for this release.

Separate LoRA download

Repository file:

fasth3_dense_datafree_6step_pruned_comfyui.safetensors

• Size: 1,036,175,488 bytes
• SHA256: 4351FEF3E48EE47970CEDD4C09DD6C1DB43DC5E96D0D8FCDE4339D37FC2418DB
• Install to: ComfyUI/models/loras/
• Recommended sampler setup for this converted file: 6 steps, model strength 1.0, CLIP strength 0.0

The publishing account holder confirmed before this release that the necessary authorization to redistribute the converted file had been obtained from the relevant rights holder. The authorization statement does not remove any license, territory, acceptable-use, notice, or downstream-recipient obligations. See LORA_PROVENANCE.md, NOTICE.md, and MINIMAX_H3_LICENSE.txt before redistribution.

Verified scope

The same-structure local production graph produced 52 completed outputs. Every file was verified as 10.125 seconds, 1728x2304, 24 FPS, H.264 video with AAC audio. The public graph was then sanitized by replacing a private local Qwen3-VL node with ComfyUI's core multiline string node. Its 43 nodes and 66 links pass structural consistency checks, but the sanitized package was not used for a new full 10-second generation after packaging. The figures therefore validate the original production path and included routing node, not universal end-to-end certification of every public installation or a universal quality claim.

Tested on 16 GB NVIDIA VRAM and 48 GB system RAM. Other configurations are not certified.

Quick start

1. Update ComfyUI to a build containing the MiniMax H3 and current VIDEO APIs.
2. Install the dependencies and place the model files listed in MODELS_AND_DEPENDENCIES_CN.md.
3. Copy custom_nodes/ComfyUI-H3-MultiFolder-Router into ComfyUI/custom_nodes/, then restart ComfyUI.
4. Open workflow/FastH3_4step_10s_batch_manual_prompt.json.
5. Enter the finished H3 video prompt in the green multiline string node. The long file under prompts is meant for a vision-language model first; do not paste that entire director instruction into H3.
6. Add matching input and output folders, one pair per line, in H3 Multi-Folder Batch Input.
7. Use the node's check button, then its queue-all button. Do not start the batch with ComfyUI's normal queue button.

Experimental settings

The workflow uses a 4-step Simple/Euler first pass at 0.4 MP, video/audio shift 12/3, learned latent upscale to 864x1152, an enabled 3-step Ref2VA pass, and optional RTX 2x output scaling.

Niko's converter documentation recommends 6 steps and warns that 4 steps can cause jitter, flicker, or color bloom. This package preserves a locally tested 4-step-first-pass plus 3-step-refinement experiment. It is not the converter author's recommended setting. Compare against 5 or 6 first-pass steps when stability matters.

MiniMax's Acceptable Use Policy requires public machine-generated content to be clearly and prominently disclosed. Confirm copyright, likeness, and guardian permission before sharing media depicting minors.

The repository-level MIT license applies only to independently authored workflow arrangement, documentation, prompt text, and router-node code. It does not apply to or relicense the separately distributed LoRA.

Prepared for the lpiao3825 community account.