This is a lossless diff-LoRA extraction of TenStrip's 10Eros_Max fine-tune of MiniMax H3, for anyone who wants the fine-tune without running the full 40 GB checkpoint — including on the ref2va (reference-to-video-audio) variant, which has no official 10Eros release.
What it is. 10Eros_Max's training touched only the fused QKV projections in transformer blocks 0–31; every other tensor is identical to the base model. This LoRA contains exactly those weight deltas, extracted by direct tensor subtraction (fine-tune minus base) and stored in ComfyUI's .diff LoRA format — not an SVD/rank approximation. At strength 1.0 on the FL2VA checkpoint it was extracted from, it reproduces the full fine-tune bit-for-bit; on sibling H3 variants it transplants the fine-tuned attention weights onto the shared transformer.
How it was made. Both checkpoints (10Eros_Max_h3_fl2va_bf16_test4_pruned and stock minimax_h3_fl2va_pruned_bf16) were streamed tensor-by-tensor; differing tensors were identified by comparison and their fp32 deltas saved as .diff / .diff_b keys in bf16. Pruned-vs-pruned extraction, so the diff matches the pruned checkpoint family's layout.
Usage.
Load with LoraLoaderModelOnly (stock ComfyUI, no custom nodes needed).
On fl2va pruned bf16: strength 1.0 = the complete fine-tune.
On ref2va pruned bf16: start at 0.7–0.8. The fine-tune was trained in FL2VA context, so at high strength it can compete with reference-image adherence — lower strength trades fine-tune character for reference fidelity. Find your balance.
Strength scales the whole fine-tune continuously — something the full checkpoint can't do.
refva turbo lora (4steps) seems to not be enough for good results yet... still testing so not 100% sure. Anyways, safer to use higher steps or simply bypass the turbo lora.
Requirements. A MiniMax H3 pruned-family checkpoint (bf16 recommended; applying onto int8/nvfp4 quants works via runtime patching but bf16 is cleanest), plus the standard H3 stack (Qwen3-VL text encoder, H3 video/audio VAEs).
Credits & license. Base model by MiniMax; 10Eros_Max fine-tune by TenStrip — all creative credit for the fine-tune is theirs, this is only a repackaging of their released weights. Use is subject to the MiniMax H3 Community License Agreement (same terms as the source checkpoints). Community conversion, not an official MiniMax or TenStrip release; TenStrip has indicated an official LoRA may accompany a future beta — prefer that when it exists.
Description
FAQ
Comments (28)
Hi, it's only pruned so ? it doesnt work with normal int 8 ?
The merge was done using 2 pruned versions. I haven't tested it yet on unpruned/normal models.
why you want non pruned version, waste of memory
Largest Size of Lora Award go to...................this
Int8 checkpoint coming out in 10 minutes. 20gb :)
12gb, at this point i just stick with eros itself lolololo
I'll see if I could quant it to make it smaller. For now this is a quick fix. Sorry for the large size :/
0.4 is work fine with int8 ref2va
Is this turbo lora?
Does this mimic the power of the raw model to pruned models?
Perhaps a stupid question but I struggle to find and understand what does the eros model do differently and how to trigger those differences?
It's meant for nsfw action (Explicit) with better movement, body anatomy etc. As it adapts some LTX properties, some normal movements may be enhanced as well
@Stuubzzz so its magic? they are different architecture right?
At this level I'm not sure if it can capture all the tiny shifts. The one I released is a full block attn layer, heads, and MLP graft to shift the entire model slightly at a lossless level. When you extract it you lose even more accuracy since all of the graft is perpendicular shaping which comes out very lossy due to SVD math. If you want to use it ref you can already do so with the int8 version with images, for videos use KJ's reference model lora extraction with it.
@tenstrip Thanks a lot for your comment!!! — to clarify, I didn't SVD this; I extracted the exact full-rank diff (18.5 GB, all the graft structure preserved) and merged it back into ref2va, so it's lossless apart from bf16 rounding. The real approximation is the fl2va→ref2va context shift, which I tuned out with strength (0.8). But I hadn't seen KJ's reference-model extraction — that sounds like it targets the ref pathway more correctly than a straight diff-merge. I'll give it a shot. Appreciate it a lot!
I actually did a couple versions and regrafted to the reference model itself, remaking the whole beta2 from the reference model. What I found was kind of troubling though, but makes sense. The graft reduces the video reference accuracy slightly in an A/B video reskin test v.s. the base ref model since it's pulling the model away with a donor model that doesn't do those per-frame references and conditioning. It still works slightly better on the images + lora kind of setup; more of a complex i2v style reference scenario with multiple images.
@tenstrip That makes a lot of sense — the orig. models never did per-frame reference, so grafting them into ref2va necessarily pulls against the frame-tracking. Good to know it's inherent and not something my diff+merge introduced. If I understood correctly: use this for multi-image / complex i2v reference composition where it's actually better, and keep base ref2va for tight video reskins where per-frame tracking matters most. Really appreciate you A/B-ing this and sharing what you found
@Stuubzzz Also I wanna point out the mistake I was making which is also in this I see. The grafted data is built on the fl model. I had to remake the grafts on the actual ref model. I put that out on huggingace already, but I have a new AIO turbo strategy I'm trying. Technically I'm not sure this would do any better ref than just using the fl model I put out, losing accuracy in the same way.
@tenstrip Sounds amazing! Can't wait to test!
I don't understand, what exactly does this LORA do? It's 12gb and doesn't replace the base checkpoint?
Some previews would be nice
It is the different of Eros and H3 normal, so you can add this to H3 normal to get Eros... and you would do that to create 3Eros on the R2V version. So you get porn + r2v.
However, FLF model can do reference if it's just a few pictures and does it with better quality since there is a bug or issue in R2V model.
@brnfd24434343d what does 3Eros do? is it train on porn it enhance base model motion?
@LuringSuccubus 10Eros/3ros/3Eros, whatever he's calling it. Yes, it's tuned on I2V porn
fl2va 4steps no work.
You have to use it with the ref2va model.
Well... without this results is better. for now.