Low-step joint video+audio generation (4 sampling steps instead of ~20) greatly reduces runtime but is fragile to model and scheduler mismatches. This repository makes the MiniMax‑H3 Turbo 4‑step LoRA usable inside ComfyUI’s pruned/curve-form MiniMax‑H3 pipeline so you can practically experiment with ultra-low-step audio‑video generation without rebuilding the full model stack.
What Sets It Apart
- ComfyUI compatibility conversion: retained tensor keys were re‑namespaced to the diffusion_model.* namespace so ComfyUI’s built‑in MiniMax‑H3 LoRA loader accepts them without key warnings.
- Partial, pruned conversions: preserved BF16 dtype, tensor shapes, LoRA ranks and A/B orientation while pruning AdaLN projection tensors so the adapters match the pruned/curve-form MiniMax‑H3 checkpoint.
- Multiple variants included: initial preview EMA/non‑EMA conversions plus further‑trained ckpt500 EMA and non‑EMA variants (time‑averaged EMA vs crisper non‑EMA behavior) to compare stability vs sharpness.
- Validation and accelerator compatibility: validated for complete A/B pairs and tested working with optional attention accelerators (SageAttention, Sol, Gradient, Spectrum) and recommended ComfyUI settings.
Key usage notes and practical guidance
- Recommended starting files: minimax_h3_turbo_4step_ckpt500_pruned_comfyui.safetensors (further‑trained non‑EMA) or its EMA counterpart for comparison.
- Recommended ComfyUI parameters used in tests: video sigma shift ~12, audio sigma shift ~4–6, sampler res_multistep, steps 8–10 (6–8 for ckpt500), LoRA strength 0.8–1.8. Audio scheduling is critical; wrong scheduler/sigma can produce destroyed or noisy audio.
- Conversion specifics: source tensors ~518 → retained 416 in pruned conversion; pruned out 102 AdaLN tensors. Namespace conversions: blocks.* → diffusion_model.blocks., token_refiner. → diffusion_model.token_refiner.*.
Who it's for and tradeoffs
Great fit if you use ComfyUI with the pruned/curve-form MiniMax‑H3 checkpoint and want to test or benchmark ultra-low-step (4‑step / 6–10 step) joint audio+video generation without rebuilding adapters. Look elsewhere if you need a full-model, unpruned Turbo LoRA or production-ready audio/video quality out of an already‑finished training run: these are partial compatibility conversions of an early preview LoRA and the original Turbo LoRA itself was an under-trained preview (EMA not fully matured).
