badincite/minimax-h3-soup

MiniMax H3 Soup

Reproducibility archive for a local ComfyUI MiniMax H3 Ref2V benchmark on an RTX 3090.

What is included

• Original benchmark workflow graph (source_prompt.json), manifest, and result table.
• Every one-second MP4 from the original C1-C11 benchmark grid and its Euler repeat sweep. The separate duration experiments are intentionally not included.
• Labeled C1-C11 visual contact sheets, Ref2VA stock-control sheets, and a static render-time summary chart.

System

• NVIDIA RTX 3090
• 32 GB DDR4 RAM
• CUDA 13
• ComfyUI 0.34.4
• PyTorch 2.9.1+cu130

Layout

Path	Contents
benchmark/primary_res_multistep_1s/	Main C1-C11 model/LoRA screen, using res_multistep at one second.
benchmark/euler_ablation_1s/	Full C1-C11 repeat grid using euler at one second.
benchmark/c7_r1_original_turbo_1s/	Corrected C7 original Larry v4 experiment using its custom loader and Turbo Sampler.
videos/primary_res_multistep_1s/	Every original-grid one-second MP4, C1 through C11.
videos/euler_ablation_1s/	The matching C1-C11 Euler videos.
videos/c7_r1_original_turbo_1s/	Corrected C7_r1 original-Turbo videos, 4/6/8 steps.
reports/primary_res_multistep_1s/	Contact sheets and the render-time chart.
workflows/	Cleaned reusable ComfyUI workflow JSON.

Dataset Viewer metadata

The primary-video folder includes a metadata.csv sidecar, so the Dataset Viewer shows a readable case (C1-C11 or Stock), display name, model/LoRA, megapixels, native resolution, steps, ksampler, seed, and measured render time beside every video. The original MP4 filenames are preserved for provenance; use the case and display name columns to browse the experiments.

Important numbering note: the original on-disk filenames use internal IDs 09_... and 10_... for the user-facing C8 and C9 tests. The metadata column maps them to the intended labels, and C10 is explicitly labelled C10.

C1-C11 model / LoRA legend

Label	Meaning
Stock	Ref2VA base with no LoRA, tested separately at 20, 25, and 30 steps
C1	minimax_h3_fused_refdelta_r1024_turbo8_mystic07_int8_convrot with no LoRA
C2	Minimax_H3_fl2va_Pruned_Lightx2v_turbo_4step_v1.2_768p_INT8 with no LoRA
C3	Minimax_H3_FL2VA_PRUNED_Turbo_8step_v1.0_768p_INT8 with no LoRA
C4	Ref2VA base + minimax_h3_fl2v_turbo_4step_v1.1_768p LoRA
C5	Ref2VA base + minimax_h3_fl2v_turbo_4step_v1.2_768p LoRA
C6	Ref2VA base + minimax_h3_ref2v_turbo_4step_v0.1 LoRA
C7 (legacy)	Ref2VA base + Larry's original minimax_h3_turbo_v4_step600_ema through the standard ComfyUI LoRA loader. Historical only; not valid for judging the original v4 LoRA.
C7_r1 (corrected)	Ref2VA base + the same LoRA through MiniMaxH3TurboLoRA and MiniMaxH3TurboSampler; simple, 4/6/8 steps, and low_vram=false sharp bypass mode.
C8	Ref2VA base + minimax_h3_fl2v_turbo_8step_v1.0 LoRA
C9	Ref2VA base + minimax_h3_fl2v_turbo_8step_v1.0_768p LoRA
C10	Ref2VA base + converted fasth3_6step_converted_dense_datafree LoRA
C11	minimax_h3_fastvideo_vsa_datafree_1300step_4step_int8_convrot with no LoRA

All primary-grid tests use a one-second duration, fixed seed, and the res_multistep KSampler. The separate Euler grid repeats C1-C11 at the same resolution/step pairs with euler; ksampler in each viewer row makes the two series directly filterable. Only resolution and step count vary within a given case. The Stock control is retained as a separate higher-step reference.

C7 correction: Larry v4 Turbo loader

The archived C7 one-second results were created before the original Larry v4 Turbo LoRA loader requirement was caught. Those legacy rows use the standard ComfyUI LoRA loader and therefore are not a valid quality ranking for the original v4 LoRA. Do not use them to judge C7 against C1-C6/C8-C11.

Use H3_REF2V_RTX3090_C7_ORIGINAL_TURBO_CUSTOM_LOADER.json instead. It uses ComfyUI-MiniMax-H3-Turbo's MiniMax-H3 Turbo LoRA node with low_vram=false (runtime bypass for maximum sharpness), MiniMax-H3 Turbo Sampler, simple scheduler, and 8 steps. A corrected C7 benchmark will be recorded as a separate experiment rather than overwriting the legacy data.

The completed corrected screen is named C7_r1 and is available in the c7_r1_original_turbo_1s Dataset Viewer configuration. The paired comparison below includes only 6 and 8 steps because those are the two step values both the legacy C7 run and C7_r1 share.

Repeatable technical-quality metrics

quality_metrics.csv and the Dataset Viewer columns include a transparent, deterministic technical score for every primary-grid video. The method is h3-transparent-v1: eight evenly-spaced decoded frames, fixed OpenCV algorithms, and fixed (absolute, not rank-based) score mappings.

The exact scoring implementation is included as tools/score_h3_primary_videos.py; it was run with OpenCV 5.0.0 and NumPy 2.4.4. Re-running it against the unchanged MP4s produces the same values.

Field	Meaning
sharpness_laplacian_variance / sharpness_score_0_100	Edge-detail proxy: higher generally means less blur, but can also reward noise or hard edges.
contrast_luma_stddev / contrast_score_0_100	Luminance contrast and a fixed score favouring a moderate cinematic range.
clipped_pixel_fraction / exposure_score_0_100	Fraction of near-black plus near-white pixels; lower clipping scores higher.
saturation_mean_0_255	Raw colour-saturation diagnostic; not part of the composite.
flow_aligned_residual_0_1 / temporal_consistency_score_0_100	Frame-to-frame residual after dense optical-flow alignment. High deliberate motion can lower this value.
technical_quality_score_0_100	45% sharpness + 25% contrast + 20% exposure + 10% temporal consistency.

These metrics do not score identity accuracy, prompt adherence, audio, motion direction, or artistic preference. They are a consistent technical filter for community review—not a replacement for watching the videos.

Visual comparison sheets

Each tile is a mid-frame from the corresponding one-second render. Columns are C1 through C11; rows are 6, 8, 10, and 12 steps. Render time is shown beneath each tile.

0.4 MP

0.6 MP

0.8 MP

0.98 MP

Render-time summary

This chart measures render speed only; it is not a visual-quality ranking.

Euler versus res_multistep

These four paired sheets compare the same C1-C11 model/LoRA, seed, resolution, and step count. Each pair is labelled RM (res_multistep) on the left and EU (euler) on the right; the time below each tile is its measured render time.

0.4 MP

0.6 MP

0.8 MP

0.98 MP

C3 versus C10 FastH3

Paired comparison at matching resolution and step counts. C3 uses the FL2VA Turbo 8-step model without a LoRA; C10 uses Ref2VA with the converted FastH3 LoRA. Prompt, seed, and duration are held constant.

results.jsonl is the append-only execution record. results.csv is the current tabular export. During the original Windows file lock, the live export was used to create the published results.csv.

Reproduction notes

The workflow JSON references local ComfyUI model paths. It does not include model weights, LoRAs, RefMods, input images, or audio assets. Install the matching MiniMax H3 nodes and models first, then update local paths as needed.

This repository is an experimental benchmark, not a claim of general model quality. Visual preference and audio quality were assessed separately; identical seed and prompt settings were used within each controlled sweep.