ethanfel/MiniMax-H3-Pruned-Ref2VA-Delta-LoRAs-Experimental

MiniMax-H3 pruned FL2VA ↔ Ref2VA delta LoRAs

Highly experimental, mechanically extracted adapters. These are randomized-SVD approximations of the weight difference between the official pruned BF16 Ref2VA and FL2VA checkpoints. They were not trained as LoRAs and have not been generation-tested. Only safetensors integrity and structural loading with ComfyUI's current LoRA parser were validated. Expect artifacts, capability loss, or unpredictable behavior.

These model-only ComfyUI LoRAs explore transferring behavior in either direction between the pruned FL2VA and Ref2VA models. The forward files approximate:

delta = pruned Ref2VA BF16 - pruned FL2VA BF16

The reverse files are exact sign-inverted forms of those approximations:

reverse delta = pruned FL2VA BF16 - pruned Ref2VA BF16

At strength 1.0, applying an adapter to its matching base moves the model toward a rank-limited approximation of the other variant. Lower strengths interpolate only in the extracted weight direction; they do not isolate a clean semantic capability.

Required checkpoints and directions

The filenames state both the checkpoint to load and the direction of the adapter. apply_to_fl2va__toward_ref2va means exactly: load the pruned FL2VA checkpoint, then apply that LoRA to move its weights toward Ref2VA.

For files named apply_to_fl2va__toward_ref2va, use this base:

Comfy-Org/MiniMax-H3
diffusion_models/minimax_h3_fl2va_pruned_bf16.safetensors

Base SHA-256:

a32572fb90b5508b201ec7c2eddcc184b13ddfd3c6f6d2cf06a0b46535d541b4

For files named apply_to_ref2va__toward_fl2va, use this base:

diffusion_models/minimax_h3_ref2va_pruned_bf16.safetensors
SHA-256: 37c0da793e20ca735272ec2be655f08a2e10f97a3ec8fdfb40f5b39a736ed6fe

Do not expect correct results with the full-size BF16, FP8, or INT8 checkpoints.

Files

Apply to this base	Move toward	File	Size
pruned FL2VA BF16	Ref2VA	minimax_h3_pruned_bf16__apply_to_fl2va__toward_ref2va__rank256.safetensors	2.589 GB
pruned FL2VA BF16	Ref2VA	minimax_h3_pruned_bf16__apply_to_fl2va__toward_ref2va__rank512.safetensors	5.075 GB
pruned FL2VA BF16	Ref2VA	minimax_h3_pruned_bf16__apply_to_fl2va__toward_ref2va__rank1024.safetensors	10.048 GB
pruned Ref2VA BF16	FL2VA	minimax_h3_pruned_bf16__apply_to_ref2va__toward_fl2va__rank256.safetensors	2.589 GB
pruned Ref2VA BF16	FL2VA	minimax_h3_pruned_bf16__apply_to_ref2va__toward_fl2va__rank512.safetensors	5.075 GB
pruned Ref2VA BF16	FL2VA	minimax_h3_pruned_bf16__apply_to_ref2va__toward_fl2va__rank1024.safetensors	10.048 GB

What rank preserves

Rank	Median capture across compressed matrices	Mean capture across compressed matrices	Global energy-weighted capture
256	14.32%	16.77%	99.1260%
512	26.40%	29.70%	99.2417%
1024	46.11%	49.62%	99.4311%

The global weighted percentage is not the full story and is not a quality score. It is dominated by a small number of high-energy tensors with low effective rank. Of the 264 matrix deltas, 54 low-rank matrices are reconstructed essentially exactly at every requested rank. The table's median and mean are calculated over the remaining 210 matrices that are genuinely compressed; typical large attention and MLP matrices retain much less energy. Rank 1024 can therefore be materially more faithful than rank 256 even though their global weighted figures differ by only about 0.3 percentage points. Weight-space energy also does not predict which directions matter most during inference; compare actual generations with identical prompts, inputs, seeds, and sampler settings.

Each .safetensors.json file contains the complete per-layer extraction report.

ComfyUI use

1. Put one .safetensors file in ComfyUI/models/loras/.
2. Follow the filename literally: load pruned FL2VA for an apply_to_fl2va file, or pruned Ref2VA for an apply_to_ref2va file.
3. Apply the adapter with a model-only LoRA loader.
4. Start at model strength 1.0; compare 0.25, 0.5, 0.75, and 1.0 using identical inputs.
5. Do not stack ranks or opposing directions together. They are alternative forms of the same underlying delta.

All six files loaded as 531 expected ComfyUI patches each: 264 matrix adapters and 267 exact vector/bias diff patches. rope.inv_freq was identical between the source checkpoints and was omitted.

Extraction details

• Matrix method: deterministic randomized SVD
• Requested ranks: 256, 512, and 1024
• Oversampling: 32
• Power iterations: 2
• Factor dtype: BF16
• Vector and bias patches: FP32 ComfyUI .diff / .diff_b
• Source tensors: 532 matching tensors
• Peak extractor CUDA allocation: approximately 1.24 GB
• Reverse-direction construction: exact sign inversion of one LoRA factor and every .diff / .diff_b tensor

The included extract_minimax_h3_ref2va_lora.py performs checkpointed, one-tensor-at-a-time extraction. It avoids loading both 40 GB checkpoints into memory simultaneously. invert_comfy_lora_direction.py creates the opposite direction without repeating or degrading the SVD.

License and territorial restrictions

These are modified MiniMax H3 model derivatives. They remain subject to the MiniMax H3 Community License Agreement and its Acceptable Use Policy. The license contains territorial restrictions, redistribution conditions, and use restrictions. Review LICENSE and NOTICE before downloading, using, or redistributing these files.

This repository does not grant additional rights or remove any upstream restriction. Users and distributors are responsible for determining whether their access, use, and redistribution are authorized.

Status

• Experimental: yes
• Trained LoRA: no
• Generation-tested: no
• Structurally parsed by ComfyUI: yes
• Affiliated with or endorsed by MiniMax or Comfy Org: no