NaukNauk/minimax-h3-ref2va-larry-v4-strength-0.7

MiniMax-H3 Ref2VA + Larry v4, strength 0.7

Premerged Ref2VA transformer only, built from the clean original model and Larry v4 Turbo LoRA at strength 0.7. This is a weight checkpoint, not a LoRA adapter or a complete pipeline.

Provenance

• Base: MiniMaxAI/MiniMax-H3@42ed227ee7df40d41602854ae760620d6eb651fe, partition Ref2VA
• LoRA: larryvrh/MiniMax-H3-Turbo-Lora@43a74557ac3f6539db8e0f2a959d03feb7a81480
• LoRA file: minimax_h3_turbo_v4_step600_ema.safetensors
• LoRA SHA-256: 5f3a626cd72c93a8b9318d6760c510bc5092d2ab13aaba1f932c5bab07a416d3
• Merge: W_base + 0.7 * (B @ A) with chunked FP32 accumulation, then conversion to each original tensor dtype
• Applied targets: 259; transformer shards: 13

The merge starts from the clean base checkpoint. It is not a rescaling of an already merged strength-1.0 checkpoint. Full provenance and file hashes are in nauknauk_turbo_merge.json.

Assembly and use

Use the transformer directory from this repository together with the non-transformer components of the pinned base model's Ref2VA partition:

<model-root>/Ref2VA/
  transformer/       # this repository
  model_index.json   # pinned MiniMax-H3 Ref2VA base
  processor/         # pinned base
  tokenizer/         # pinned base
  text_encoder/      # pinned base
  video_vae/         # pinned base
  audio_vae/         # pinned base

Serve the assembled root with MiniMaxH3Pipeline and --model-variant ref2va. Preserve any other non-transformer files in the base partition. Do not apply the Larry LoRA again on top of this merged transformer.

Validation scope

All 259 LoRA target tensors were checked against the live model across TP=2 and Ulysses=2 on four H100 GPUs. A reference-image-to-video smoke test completed with video and audio using SGLang 0.5.17, a 4-second request, short edge 480, aspect ratio 16:9, num_inference_steps=10, tail extension 0, video shift 12 and audio shift 3. Details are in validation.json.

These checks establish merge integrity and successful inference. They are not a quality comparison or a recommended optimum for every prompt, resolution, or step schedule.

License and access

This derivative remains subject to the MiniMax-H3 community license in LICENSE and the upstream model's terms. The source LoRA declares license apache-2.0; its license is copied as LICENSE-LORA when provided upstream. Credit: MiniMax and larryvrh.

Access is manually gated, matching the existing NaukNauk MiniMax-H3 merged transformer repository.