beta 3:
Rebuilt on the delta1024 reference hybrid model. Does t2va and reference. Treat i2v as a single image reference, don't do i2v prompting. Silver's merged turbo is integrated and tuned for 6 steps with simple schedule with the extra _emb layers tacked on to the model, not sure if they're needed.
Samplers like er_sde or multires or other turbo sampling setups work. Best imo is just er_sde/simple 6 steps, no shift, no spectrum, no cache, only a comfy_kitchen backend selection node. 3 video references in a 15+ second outputs can be done in under 10-15 minutes now on larger cards with no extra cache or quality hit needed.
Many (like a lot, all the good ones) on-site loras were fully combined into a consensus weighted merge with ranked drop-out to form an initial part. That merge is put against new, more powerful Wan and LTX grafts as a blend/reshape that uses the loras to consensus shape the grafts, but it also allowed some of the better loras clean pass-through. This is not a linear list of loras just merged. The main element that can present most is probably MysticXXX which was given the most pass-through weight since it's just good--and all 3 release steps of that are inside it. However, they're all-combined with a ton of other loras with agreement and consensus of shape and then it's only reshaping the grafts parts. The results of that are the actual weighted loras that are used to make the model.
Due to drop-out and consensus merge, pretty much all of the loras can all still be used easily on top if needed, and might work better even. All of this was only done to create a large rank dummy lora similar to what sulphur data will look like as a lora or extracted lora so I can start looking at how to apply it cleanly.
It's definitely not a few on-site loras that are linear merged, uncredited, and then renamed with some emojis. I'll only do this until sulphur tuning steps are in my hands and I can work with more targeted and shifted stuff, plus I was tired of waiting and I wanted fast easy i2v.
There is one quirk of the hybrid h3 usage: don't use it i2v. It should either be always used in reference prompting mode, or t2va prompting mode. Even if there is just one single image input it needs to be used as reference and prompted in the ref2va format. If you run an underdeveloped or manually written prompt you will get odd outputs, random camera changes, and blue lighting color shifts when you use the i2v prompt style.
Full credits to these lora makers for being involved somewhat in beta3 version:
alcaitiff, MisticRain69, diogod, FourBunny, HearmemanAI, tazmannner, simonishere, QualityControl, blo01, ComfyTinker, kermitfrog1202(Playtime)
Beta2 and previous:
This started a finetune-by-graft. Or maybe a GST - grafted shift of transformer (cross-architecture). I made both up, because there aren't any projects that have done it that I know, except one reddit post that made me look into it. I experimented with Wan and LTX on the side which led to the initial LTX Eros scripts that became what powered this, all before H3 ever came out. It seems like unified unbiased models like MMH3 can technically take attention influence from any other DiT without breaking if done correctly. Anima, Krea2, LTX, Wan2.2, Flux1 were all tried out, configs tested, about ~40 hours maybe of working in the dark without any paper or technical documents from Minimax. Eventually I developed linear-magnitude blend application and specific block and head gate targets allowing for a smoother graft on an attn-triplet-unfused version of H3 output as a patch file. That sent to lora extraction, then merged to checkpoint at taste. This is a merge but a merge of LoRas I extracted that interact to produce this current shift. I saved 5 ponds of water by recycling data in a few minutes on a single card instead of toasting a server up.
Turbo not recommended yet for i2v, especially when used with other LoRas. T2V use with turbo is better. Use 20-25 steps normal sampling with no dialogue, 25 steps with dialogue along with cache nodes and attn modes. More steps over 25 are not neccessarily better, and can be worse. Use full int8: int8 model, int8 VAE (if it doesn't crash comfy), int8 qwen3vl along with current cache or attn mode nodes. For smaller cards: quants, macOS ports, and Wan2gp support will likely appear on huggingface but not from me.
Known quirks:
Audio difference v.s. Base - This model's audio changes come from attention shifts seeking alternate audio pairing. Attn triplets were unfused before graft, both standard and triplet q_attn was grafted holding about maybe 10-15% audio influence, attn_k was frozen and MLP fc2 layers were untouched resulting in minimal audio interference. This was the main issue with the entire transformer graft and protecting audio. However this version is slightly louder overall than the base model.
Low resolution detail smearing - Some finger digits and fine motion will smear more at low resolution, also a problem in base model. As memory use gets more efficient increase resolution or work on the composition to get around it.
Odd outputs - This can attempt certain concepts more liberally than base model, but that can lead to some undesirable outputs in bad prompting and certain contexts. Data shift comes from completely different transformers and architecture. This shouldn't even work, so it is what it is.
This model is not dedicated to NSFW as that would violate community license agreement. Sure it can do it, just like base. Any NSFW generations are purely the result of advanced reasoning and tokenization resulting from experimental changes. All terms from the H3 community license also still apply to the users of this version. Don't be a dumbass.
H3 usage still requires very intense prompting for maximum effect. Every motion, every interaction, every sound plainly and fully described. Not with slang terms; with proper actionable words that can be tokenized. Refer to the h3 developer prompting guide, hand that .md file to an LLM or Chat agent and have them enhance or refine prompts along the released H3 developer prompt guide styles using the model's tag system. Certain concepts can be made from pure token reasoning. Consult the prompts in my previews to see certain physical descriptions that I use for some things. When using enhancement give the agent feedback about any issues in the generation and get them to describe motions in alternate fashion, or manually edit it yourself adding a negative like "no X, no Y". Still requires prompt refinement and trial/error for best outcomes.
Sulphur Project has 10k banked to attempt actual tuning. Right now training pipelines are sub-optimal. As always Eros is my personal side project, and this beta was also essentially a speed-run of finetuning, figuring out exactly in what configurations and target areas do you get helpful/harmful changes in the model. This is also a proof-of-concept of what and where to target while leaving the reinforcement quality of base unharmed by being additive.
Description
Hybrid delta base, turbo inject, graft+loras drop-out rescale, and consensus TIES merge.