--- license: apache-2.0 base_model: Comfy-Org/MiniMax-H3 base_model_relation: adapter pipeline_tag: text-to-video tags: - minimax-h3 - lora - adapter - comfyui - pruned - pruned-model - curve-form - text-to-video - text-to-audio - audio-video - video-generation - audio-generation - synchronized-audio - four-step - 4-step - turbo - accelerated-inference - further-trained - checkpoint-500 - safetensors - bfloat16 - bf16 - partial-conversion --- # MiniMax-H3 Turbo 4-Step — ComfyUI Pruned-Model LoRAs This repository contains third-party ComfyUI compatibility conversions of the original **MiniMax-H3 Turbo LoRA — 4-step audio-video generation preview** created by **[larryvrh](https://huggingface.co/larryvrh)**. It also includes two **further-trained checkpoint-500 variants** in the same pruned-model ComfyUI format. These LoRAs are intended for the **pruned/curve-form MiniMax-H3 checkpoint used by ComfyUI**. **also check https://github.com/Larryvrh/ComfyUI-MiniMax-H3-Turbo** ## ComfyUI Test Output [Download the ComfyUI workflow](https://huggingface.co/drbaph/MiniMax-H3-Turbo-Lora-ComfyUI/resolve/main/fl_minimax_h3_turbo_lora_example_workflow.json?download=true) ![ComfyUI workflow screenshot](https://cdn-uploads.huggingface.co/production/uploads/63473b59e5c0717e6737b872/zfVV_yKP-njnOrGinGy8K.png) ## Original Project The original LoRA weights, training work, four-step distillation method, dual video/audio sampling implementation, documentation and source files were created and released by **larryvrh**: **Original repository:** [larryvrh/MiniMax-H3-Turbo-Lora](https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora) I did **not** train, distill or create the original Turbo LoRA weights. This repository provides modified compatibility versions intended to allow the compatible LoRA adapters to load with ComfyUI's built-in MiniMax-H3 LoRA loader when using the pruned/curve-form model. Full credit for the original MiniMax-H3 Turbo LoRA and its distillation method belongs to **larryvrh**. ## Included Files | File | Description | |---|---| | `minimax_h3_turbo_4step_pruned_comfyui.safetensors` | Initial partial ComfyUI compatibility conversion of the original non-EMA Turbo weights for the pruned/curve-form MiniMax-H3 checkpoint. The original author describes the non-EMA weights as crisper and better at holding fast motion. | | `minimax_h3_turbo_4step_ema_pruned_comfyui.safetensors` | Initial partial ComfyUI compatibility conversion of the original EMA Turbo weights for the pruned/curve-form MiniMax-H3 checkpoint. The original author describes the early EMA weights as smoother but softer because the EMA had not fully matured. | | `minimax_h3_turbo_4step_ckpt500_pruned_comfyui.safetensors` | Further-trained checkpoint-500 non-EMA variant in the pruned-model ComfyUI format. This continues training beyond the initial preview checkpoint. | | `minimax_h3_turbo_4step_ema_ckpt500_pruned_comfyui.safetensors` | Further-trained checkpoint-500 EMA variant in the pruned-model ComfyUI format. This contains the time-averaged weights from the further-trained checkpoint. | | `fl_minimax_h3_turbo_lora_example_workflow.json` | First-and-last-frame ComfyUI example workflow using the Turbo LoRA, with optional attention and memory optimizations. | The EMA and non-EMA files remain separate and contain different tensor values. The original converted files preserve the initial preview weights, while the `ckpt500` files provide the further-trained variants. These are partial compatibility versions of the original full-model Turbo LoRAs for use with the pruned/curve-form MiniMax-H3 architecture. ## Which File Should I Use? For the further-trained weights, start with: `minimax_h3_turbo_4step_ckpt500_pruned_comfyui.safetensors` Use the further-trained EMA variant for comparison: `minimax_h3_turbo_4step_ema_ckpt500_pruned_comfyui.safetensors` The two files without `ckpt500` preserve the earlier preview versions: - `minimax_h3_turbo_4step_pruned_comfyui.safetensors` - `minimax_h3_turbo_4step_ema_pruned_comfyui.safetensors` The non-EMA variants generally preserve a sharper response and stronger fast-motion behavior, while the EMA variants are intended to provide smoother time-averaged weights. ## About the Original Turbo LoRA The original MiniMax-H3 Turbo LoRA is an early preview checkpoint designed to generate joint video and synchronized stereo audio in approximately **4 sampling steps instead of the usual approximately 20 steps**, providing roughly a **5× reduction in sampling wall-clock time**. The original author clearly states that the initial release is an unfinished preview checkpoint from an in-progress training run: - it is under-trained; - the initial EMA weights have not fully matured; - quality is not representative of a completed run; - it is intended as an early demonstration of the direction of the project. The additional `ckpt500` files in this repository are further-trained variants beyond that initial preview checkpoint. Refer to the original repository for the original author's documentation, generation script, scheduler implementation and future updates: [larryvrh/MiniMax-H3-Turbo-Lora](https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora) ## Why This Conversion Was Made The original Turbo LoRA files use tensor keys such as: `blocks.0.attn.qkv_proj.lora_A.weight` ComfyUI's built-in MiniMax-H3 LoRA resolver expects keys under the model namespace: `diffusion_model.blocks.0.attn.qkv_proj.lora_A.weight` Without that namespace, ComfyUI reports warnings such as: `[WARNING] lora key not loaded: blocks.0.attn.out_proj.lora_A.weight` The following namespace conversion was applied to the retained tensors: - `blocks.*` → `diffusion_model.blocks.*` - `token_refiner.*` → `diffusion_model.token_refiner.*` The retained tensor values, BF16 dtype, tensor shapes, LoRA rank and A/B orientation were preserved. ## Conversion Details Each original full-model Turbo LoRA contained: | Component | Count | |---|---:| | Source tensors | 518 | | Source LoRA A/B pairs | 259 | Each pruned-model compatibility conversion contains: | Component | Count | |---|---:| | Retained tensors | 416 | | Retained LoRA A/B pairs | 208 | | Pruned AdaLN tensors | 102 | | Pruned AdaLN A/B pairs | 51 | The pruned adapters were: - 50 AdaLN projection pairs from `blocks.0` through `blocks.49`; - 1 AdaLN projection pair from `final_layer.adaln_proj.linear`. The retained adapters include: - all main-block attention adapters; - all main-block MLP adapters; - all token-refiner attention adapters; - all token-refiner MLP adapters. The retained attention, MLP and token-refiner tensors have dimensions compatible with the pruned/curve-form MiniMax-H3 checkpoint. ## Validation The converted LoRA structure was checked for: - complete LoRA A/B adapter pairs; - no orphan A or B tensors; - retained keys using the `diffusion_model.*` namespace; - no duplicated `diffusion_model.diffusion_model.*` prefixes; - no incompatible `adaln_proj` tensors; - preserved BF16 dtype; - preserved tensor shapes; - preserved LoRA ranks; - separate EMA and non-EMA weights; - no unresolved retained keys during ComfyUI loader-level validation. The original source files were not overwritten. The LoRAs have been tested working in ComfyUI with the following optional acceleration methods: - **SageAttention** - **Sol Attention** - **Gradient** - **Spectrum** These acceleration methods are optional and are not included in this repository. ## Important Limitation These files are **partial ComfyUI compatibility versions for the pruned/curve-form MiniMax-H3 checkpoint**. The further-trained `ckpt500` variants improve upon the initial compatibility versions, ## Usage Place the LoRA files inside: `ComfyUI/models/loras/` Use them with the **pruned/curve-form MiniMax-H3 checkpoint for ComfyUI**. Recommended further-trained version: `minimax_h3_turbo_4step_ckpt500_pruned_comfyui.safetensors` Further-trained EMA version: `minimax_h3_turbo_4step_ema_ckpt500_pruned_comfyui.safetensors` Suggested ComfyUI settings: - Video sigma shift: `12` - Audio sigma shift: `4-6` - Steps tested working: `8–10` / `6-8` (ckpt500) - Recommended sampler: `res_multistep` - Suggested LoRA strength: `0.8–1.8` - A higher LoRA strength can generally be used with fewer sampling steps - Use the same prompt, seed, resolution, model checkpoint, sampler and workflow when comparing files Optional accelerators confirmed working: - **SageAttention** - **Sol Attention** - **Gradient** - **Spectrum** Do not assume broken audio is caused by the LoRA before checking the sampler, scheduler and audio sigma-shift configuration. ## Audio and Scheduler Notes MiniMax-H3 generates video and synchronized audio as separate streams with different scheduling requirements. The original author's standalone generation implementation uses a dual video/audio scheduling system specifically designed for ultra-low-step generation. For this ComfyUI pruned-model version, the tested settings are: - Video sigma shift: `12` - Audio sigma shift: `4-6` - Steps: `8–10` / `6-8` (ckpt500) - Sampler: `res_multistep` Incorrect audio scheduling may result in: - blown-out audio; - severe distortion; - noise-like audio; - unstable audio energy; - completely broken audio. For the original full-model settings, model files, resolution guidance, frame-count rules and dual-scheduler implementation, follow the original README: [larryvrh/MiniMax-H3-Turbo-Lora](https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora) ## Attribution Original MiniMax-H3 Turbo LoRA author: **[larryvrh](https://huggingface.co/larryvrh)** Original project: **[MiniMax-H3-Turbo-Lora](https://huggingface.co/larryvrh/MiniMax-H3-Turbo-Lora)** This repository contains third-party, partial ComfyUI compatibility versions for the pruned/curve-form MiniMax-H3 checkpoint, including further-trained checkpoint-500 variants. I do not claim authorship of: - the original Turbo LoRA weights; - the original LoRA training; - the original distillation method; - the original four-step sampling method; - the original dual video/audio scheduler; - the original `generate.py`; - the original documentation; - the MiniMax-H3 base model; - the pruned/curve-form MiniMax-H3 checkpoint. Please retain attribution to **larryvrh** and comply with the licences and terms of the original LoRA and MiniMax-H3 base model when using or redistributing these files.