Stuubs/10eros_Max_Ref2va

---

license: other license_name: minimax-h3-community-license license_link: LICENSE tags: - video - minimax-h3 - ref2va - int8 - bf16 - image-to-video base_model: - TenStrip/10Eros-Max - MiniMaxAI/MiniMax-H3

10Eros_Max beta2 — MiniMax H3 ref2va

Community builds of TenStrip's 10Eros_Max (beta2) fine-tune on the MiniMax H3 ref2va (reference-to-video-audio) checkpoint, in both bf16 and native int8_convrot.

All creative credit for the 10Eros_Max fine-tune belongs to TenStrip. This repository provides ref2va variants and int8 quantizations of their released weights. No new training was performed.

Two approaches in this repo

TenStrip's 10Eros_Max fine-tune is grafted onto the FL2VA checkpoint. Bringing it to ref2va can be done two ways, and both are provided here for comparison:

1. Native ref2va — TenStrip's own regraft, built directly on the reference model from scratch. This is their official ref2va release, quantized here to int8.
2. Transplant — the FL2VA fine-tune's exact weight changes isolated as a diff and applied to the stock ref2va checkpoint at strength 0.80. A community approximation, made before the native version was released.

Per TenStrip's own testing, both approaches slightly reduce per-frame video-reference accuracy versus the plain base ref2va model — because 10Eros_Max grafts patterns from donor models (LTX, Wan, Krea) that have no per-frame reference conditioning, so grafting them into a reference model intrinsically dilutes reference tracking. This is inherent to the concept, not a quantization or transfer artifact. Both variants shine in multi-image / complex I2V-style reference composition; for tight video-reskin / per-frame reference tracking, the plain base ref2va tracks better.

Files

File	Approach	Format	Size	Notes
10Eros_Max_h3_ref2va_beta2_native_int8_convrot.safetensors	Native regraft	int8_convrot	~21 GB	Recommended — TenStrip's native ref2va
10Eros_Max_beta2_s080_h3_ref2va_pruned_int8_convrot.safetensors	Transplant s0.80	int8_convrot	~21 GB	Community diff-transfer
10Eros_Max_beta2_s080_h3_ref2va_pruned_bf16.safetensors	Transplant s0.80	bf16	~40 GB	Full-precision transplant

If you just want the best ref2va 10Eros, use the native int8. The transplant builds are kept for comparison and for anyone wanting the strength-tunable diff approach.

Quantization fidelity (int8 files)

Both int8 files were quantized with ComfyUI's own TensorWiseINT8Layout convrot quantizer (group-wise Hadamard rotation, groupsize 256), using the official ref2va int8_convrot file as a structural template — all comfy_quant configuration blobs and unchanged tensors are byte-identical to the official release, so they load through the exact same native int8 path. Calibration against the official file: 100% of elements within ±1 quantization step. Expect ordinary int8-vs-bf16 differences and nothing more.

Usage

Load like any H3 ref2va checkpoint — same loaders, no LoRA node, no custom nodes. Requires the standard H3 stack (H3-truncated Qwen3-VL text encoder, MiniMax video + audio VAEs) and a ComfyUI recent enough for native comfy_quant int8_convrot loading. Ref2va prompting uses the six-section Full-Reference format from the MiniMax-H3 repo.

Provenance

• Native ref2va: TenStrip's 10Eros_Max_h3_ref2va_beta2_pruned, quantized here.
• Transplant: 10Eros_Max_h3_fl2va_beta2_pruned diffed against minimax_h3_fl2va_pruned_bf16, applied to minimax_h3_ref2va_pruned_bf16 at strength 0.80.
• int8 quantization template: the official minimax_h3_ref2va_pruned_int8_convrot.

Notes & limitations

• beta2 is an active TenStrip experiment and may be superseded by later betas — check their repo.
• Reference-accuracy trade-off applies to both variants (see "Two approaches" above).
• Community observation on the 10Eros lineage: check audio output on your own content.

Credits & license

• Base model: MiniMax — MiniMax-H3
• Fine-tune + native ref2va regraft: TenStrip — 10Eros-Max — all creative credit theirs
• int8 quantization + transplant build: community work, not affiliated with or endorsed by MiniMax or TenStrip

Governed by the MiniMax H3 Community License (same terms as the source checkpoints). These are Model Derivatives under that agreement; the license text is included in this repo as LICENSE.

More info on readme