Abiray/10Eros-Max-ref2va-Beta2-GGUF

10Eros-Max-Ref2VA-Beta2-Pruned (GGUF)

This repository contains heavily optimized GGUF quantizations of [TenStrip/10Eros-Max](https://huggingface.co/TenStrip/10Eros-Max) (specifically the 10Eros_Max_h3_ref2va_beta2_pruned version).

These models are designed to run in ComfyUI natively via ComfyUI-GGUF, allowing users with consumer GPUs (12GB - 16GB VRAM) to run this massive 40GB+ MiniMax-H3 model locally. The ref2va version features TenStrip's specific grafting recipe combining weights from LTX, Wan, and Krea.

📦 Available Quantizations

🚀 How to Use in ComfyUI

1. 1.Install Custom Node: Install ComfyUI-GGUF (by city96) via the ComfyUI Manager.
2. 2.Download Model: Download your preferred .gguf file (Q4KM is recommended).
3. 3.Placement: Place the downloaded .gguf file inside your ComfyUI/models/unet directory.
4. 4.Workflow: Replace your standard Load Diffusion Model node with the `Unet Loader (GGUF)` node.

⚙️ Recommended Inference Settings

Note: This is the Standard / Pruned Beta 2 version, NOT the Turbo version.

• —Steps: 20 to 30
• —CFG Scale: ~3.5 to 4.5
• —Sampler/Scheduler: Standard MiniMax flow matching settings (e.g., euler / simple or flowmatch)

📜 Original Model Information & Credits

All credit for the underlying model architecture, grafting methodology, and training goes to TenStrip and the MiniMax team.

From the Original Creator (TenStrip):

For TURBO model use: multires/simple 6 steps, ersde/simple 6 steps, or I prefer and merged it for ersde with custom 7 step sigmas to get rid of motion noise details. Er_sde/simple and custom sigma string instead of basic scheduler: [1.00, 0.94, 0.83, 0.72, 0.55, 0.30, 0.10, 0.00]

Turbo from silveroxides/MiniMax-H3_tests/tree/main with silu(t_emb) grid restored before merge.

Due to a ton of confusion I've let Claude compile a full MD on the grafting, including code and methodology as it was mostly used to create code, document, and manage the graft project while I tested, architected, and mixed. I was keeping that to myself since maybe it'd actually be nice to have something to myself but after a handful of ignorant comments I'm open sourcing everything from it except the scripts themselves. You can hand the md back to claude or any competent agent and graft in the same way or have them explain it until you understand. No more dumbassery allowed now.

This is an evolving project subject to future fixes in H3 training. Training is clearly problematic, and my branch will rely on Sulphur H3 tunes. But I didn't want to wait for all that so I took the data from older models, grafting it in where the model needs it, grafting it to attn layers at a low level that doesn't disturb H3's visual or audio output quality. That's essentialy it.

License

The standard H3 community license applies (minimax-h3-community-license-agreement). Because this release now carries transferred character from LTX 2.3, Wan 2.2, and Krea 2, the community licenses for those source models apply as well to the portions of character that came from each.