[!CAUTION]
This is an experimental weight modification, not an official MiniMax release and not a proven
“de-oil”, “de-wax”, restoration, or quality-fix model. The original checkpoint remains the
recommended default.
[!IMPORTANT]
MiniMax H3 is governed by the
MiniMax H3 Community License Agreement,
which defines excluded territories and mandatory redistribution conditions. Before publishing or
redistributing this derivative, include the official LICENSE, keep the modification notice and
NOTICE, and confirm that the intended distribution method and audience are authorized. A public
Hugging Face repository may be reachable from excluded territories; a repository gate alone is not
necessarily geographic access control. This model card is not legal advice.
Native INT8 ConvRot container inherited from the source repack
Training
None
Modified tensor
video_patch_proj.weight only
Input patch HF gain
1.02
Output-head gain
1.00 — unchanged
File size
34,038,894,550 bytes, approximately 31.70 GiB
Status
Experimental A/B candidate
What was changed
MiniMax H3 video latents are patchified with a 1 × 2 × 2 patch. For every latent channel, the four
spatial input columns of video_patch_proj.weight were transformed in an orthonormal 2×2 Haar basis:
the DC/common component is preserved at gain 1.00;
the three non-DC H/V/D components are multiplied by 1.02;
the result is transformed back and written into the input projection;
no output head, audio tensor, shared Transformer block, VAE, text encoder, or ComfyUI node was changed.
The equivalent 4×4 transform has diagonal 1.015 and off-diagonal -0.005. Its all-ones/DC
eigenvector has gain 1.00; the three orthogonal spatial-detail eigenvectors have gain 1.02.
This is not an output sharpening filter. It changes the video latent input projection used during
each joint audio-video denoising forward pass. It cannot reconstruct real texture that the model does
not generate.
DC-sum maximum absolute error: 9.5367431640625e-07;
SHA-256 of all unchanged file regions:
82d4866debf06c9b9b7a67d645d1dcac57bdd91eba3b341a55a4e1fe9f687852.
The machine-readable surgery record is available at evidence/surgery.json.
ComfyUI usage
Required model components
This repository contains only the modified Ref2VA diffusion model. Obtain the remaining H3 components
from the official or ComfyUI repack repositories under their applicable licenses:
MiniMax H3-compatible Qwen3-VL text encoder;
minimax_h3_video_vae_fp16.safetensors;
minimax_h3_audio_vae_fp32.safetensors;
a current ComfyUI build with MiniMax H3 and native INT8 ConvRot support.
Load the checkpoint with ComfyUI's Load Diffusion Model node and use a Ref2VA/R2V workflow. The
official ComfyUI template is available at
video_minimax_h3_r2v.json.
Recommended comparison protocol
For a meaningful A/B, change only the diffusion model:
generate once with minimax_h3_ref2va_int8_convrot.safetensors;
generate again with minimax_h3_ref2va_patchin_hf102.safetensors;
keep seed, reference media, prompt, dimensions, frame count, sampler, scheduler, shifts, steps, VAEs,
text encoder and conditioning identical;
do not add a Turbo/four-step LoRA when reproducing the validation below;
The local validation used the isolated
minimax-h3-audio-T8 ComfyUI nodes with
dual_clock_euler + native_flow, video shift 12, audio shift 3, and 20 joint denoising steps. That
custom node package is the validation environment, not an architectural requirement of the checkpoint.
The 22-frame test is a short diagnostic, below the official 4–15 second usage range. It does not prove
quality at the recommended 124+ frame range. One centered portrait reference and one prompt were used;
the reference image is not redistributed.
The original checkpoint was generated twice under seed 2608103502. Decoded video was pixel-identical
and decoded PCM audio was sample-identical, confirming deterministic repeatability for this setup.
Candidate-to-source ratios
Seed
Skin gradient
Skin Laplacian
Skin high-pass
Positive skin ROIs
Temporal change
Audio correlation
Audio RMS
2608103502
1.0325
1.0760
1.0189
2/3
1.0321
0.9930
0.9189
2608103503
1.0266
1.0418
1.0131
3/3
1.0134
0.9687
1.0154
The skin high-pass proxy increased by approximately 1.89% and 1.31% across the two seeds, without a
fixed 2×2 parity artifact. This is a weak proxy-level signal only. Manual inspection did not establish
a convincing reduction of oily or waxy skin appearance.
The first seed also showed an approximately 8.1% audio RMS reduction. H3 predicts audio and video in a
joint denoising trajectory, so changing a video-only input projection can still indirectly change the
generated audio.
Full machine-readable results:
evidence/metrics_seed_2608103502.json
evidence/metrics_seed_2608103503.json
Visual comparisons
Each comparison is source / input HF 1.01 / input HF 1.02 from left to right. The 1.01 candidate is
shown only as an additional control and is not included in this repository.
Seed 2608103502:
Seed 2608103502 comparison
Seed 2608103503:
Seed 2608103503 comparison
Limitations
This is not an official MiniMax model.
It is not a trained fine-tune, LoRA, distillation, restoration model or calibrated de-wax model.
Only one tensor was heuristically transformed; the direction was not learned from paired data.
Evaluation covers two seeds, one portrait reference, one prompt and a short 22-frame diagnostic.
No official-range 124–362 frame matrix, 768p/2K test, multi-subject test, blind preference study or
broad prompt evaluation has been completed.
Small proxy-metric gains may not correspond to perceptual improvement.
Identity, framing, motion and generated audio may change even when the modified tensor is video-specific.
Compatibility with Turbo LoRAs, third-party samplers, alternate quantization loaders and older ComfyUI
versions is not established.
The source checkpoint remains the recommended default for production or reproducible comparisons.
Intended use
Intended only for controlled research and A/B evaluation of a small, auditable MiniMax H3 video-input
projection change, subject to the upstream license and acceptable-use policy.
Not intended for claims of guaranteed quality improvement, identity preservation, medical/forensic use,
deceptive impersonation, bypassing safety systems, or any use prohibited by the MiniMax H3 license.
License and attribution
MiniMax H3 and this Model Derivative are governed by the
MiniMax H3 Community License Agreement.
The upstream license, including territorial restrictions, redistribution requirements, commercial terms
and acceptable-use policy, takes precedence over this model card.
Before distributing the checkpoint:
include an unmodified copy of the official MiniMax H3 LICENSE in the repository;
include the provided NOTICE;
retain a prominent statement that minimax_h3_ref2va_patchin_hf102.safetensors is modified;
comply with the applicable-territory and downstream-user requirements;
verify any commercial-use authorization requirements directly with MiniMax.
minimax_h3_ref2va_patchin_hf102.safetensors is a modified file derived from the MiniMax H3 Ref2VA
checkpoint. It was created without training by applying a 1.02 non-DC 2×2 Haar gain to
video_patch_proj.weight. This modification is independent and is not endorsed by MiniMax or Comfy-Org.
