10Eros_Max beta2 × MiniMax H3 ref2va — INT8 ConvRot
An int8_convrot quantization of TenStrip's 10Eros_Max work on the MiniMax H3 ref2va (reference-to-video-audio) variant, at less than a third of the bf16 size. Drop-in for standard H3 ref2va workflows: same loaders, same native int8 kernels as the official Comfy-Org int8 files, no LoRA node, no custom nodes.
What it is (per TenStrip, the author). This build is Wan VACE layers on H3 out_proj, plus Krea2 turbo weights across the mid-late blocks, on top of the beta2 Wan attention graft — then merged with Silver's turbo and the timetable layer restoration. It is an early experimental turbo merge (a first test build at the time of quantizing), not a finished release.
Intended use. Primarily for video reference and fast multi-video reference — this is a turbo/video-oriented model. For pure image reference, TenStrip's FL model performs better and is the recommended choice. Match the model to the task: this one for video/multi-video reference scenarios, the FL model for image reference.
Quantization fidelity. Quantized with ComfyUI's own TensorWiseINT8Layout convrot quantizer (group-wise Hadamard rotation, groupsize 256, deterministic rounding) — the same code path behind the official int8_convrot releases — using the official ref2va int8 file as a structural template. Calibration against that official file: 100% of elements within ±1 quantization step; all comfy_quant configs and unchanged tensors byte-identical to the official release. Expect ordinary int8-vs-bf16 differences, nothing more.
Requirements. 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 (if the official H3 int8 files load for you, this loads). Ref2va prompting uses the official six-section Full-Reference format — see the MiniMax-H3 repo's ref-mode prompt guide.
Credits & license. Base model by MiniMax; all 10Eros_Max fine-tune and graft work by TenStrip — all creative credit is theirs. This is a community-made quantization of their released weights, not an official MiniMax or TenStrip release. Use is governed by the MiniMax H3 Community License Agreement (same terms as the source checkpoints — license linked from the base model repo).
Description
v1.0 — beta2 · ref2va · s0.80 · int8_convrot (recommended)
First release. TenStrip 10Eros_Max beta2 (2026-08-15 upload) transferred to MiniMax H3 ref2va, baked at strength 0.80, native int8_convrot (~21 GB).
Built from:
10Eros_Max_h3_fl2va_beta2_pruned+minimax_h3_ref2va_pruned_bf16; quantization template: officialminimax_h3_ref2va_pruned_int8_convrot.90 tensors changed vs stock (blocks 0–29: qkv_proj, out_proj, mlp.fc1); everything else byte-identical to the official int8, including all
comfy_quantconfigs.Strength 0.80 chosen after full-strength transfer showed occasional artifacting on ref2va; tested clean at 0.80.
Quantized with ComfyUI's own convrot int8 quantizer; calibration vs the official file: 100% of elements within ±1 step.
Known limitations: beta2 is an active TenStrip experiment — expect this version to be superseded when beta3 lands. Cross-variant transfer: trained in FL2VA context; reference adherence tested working, but heavy multi-subject scenes may behave slightly differently than stock. Community note from the 10Eros quant ecosystem: judge audio output on your own content.
FAQ
Comments (16)
Thanks for your hard work! Can't wait to try it out.
I also created a bf16 version so let me know if you want to test that out too. For me the generation times with bf16 were around 50% higher than int8 so figured I might as well just release the int8 for now.
I don't get it? do I put this this in the diffusion modelds folder or lora folder?
I put it in the diff folder but there is no way to adjust the strenght
diff model is right.I thk he used 0.8 sth to trained 10eros max modle.
It's a full diffusion model. It was created by merging the Minimax int8 with a lora pulled from 10eros. So no need for a lora. The 10eros was merged with strength of 0.8 as when I tried a merge with 1.0, I started getting some artifacts so I lowered it to 0.8.
@drfaker911219 I created the 10eros also as a lora. You can adjust the strength if you'd rather use the lora instead of the full diffusion model. It's a quite beefy lora though (12gb) as I haven't had the time to quant it yet :/
Somehow i am not able to run it ...
UnicodeDecodeError: 'utf-8' codec can't decode byte 0xe9 in position 95617: invalid continuation byte
Hi! Sorry to hear it didn't load properly. I will check if something in the script caused the issue. Meanwhile verify the downloaded file's size/hash matches Civitai's, re-download if it doesn't. Does the normal comfyui int8 convrot work for you? (the one you can download from Huggingface "Minimax Comfy")
@Stuubzzz using minimax int8 convrot without any issue... I have checked size download seems to be ok... Also tried more workflows but still same error
@fabiox3 Got it. I'll check it later and will change the model if something is wrong with the script that was used for baking the merge. I hope we get it working for you :)
@Stuubzzz Works here along with other models without an issue. Not as stable though when comparing to Xueluo ref model for example.
@fabiox3 new int8 released. Tenstrip also released his own bf16 version which you could give a go :) Hope things work now
I can't get a perfect anatomy by putting only images of character references and prompt the sex. Can it make a correct anatomy this way? Or always have to do a First frame with anatomy exposed?
I had the same issue before, but I sorted it out by having another reference image purely of the parts of the anatomy that were incorrect. As the 10eros progresses, I'm sure the anatomy parts and some actions will become better after additional training of the model. But for now your best bet is to use specific images in addition to your character image to enhance the body parts / anatomy.
Be sure you have the right structured ref prompt. <Subject 2>'s genitals are fully referenced in <Picture 1>, pussy and breasts. Like that.