Stubelius Remix v1 (LTX 2.5)
The remix is out of beta. v1 is beta 2 plus one thing I tested properly before baking it in: OmniNFT, a reinforcement-learning fine-tune that rewards video and audio quality, merged in at its trained strength. Everything else is the same as beta 2. It’s still mainly meant for NSFW work but can do SFW (name the clothing in SFW prompts, or you may get unwanted nudity), and the distilled LoRA is still baked in.
What’s new (measured against beta 2 on my test prompts)
Sharper fine detail: +55%. Faces, hair, fabric and backgrounds come out crisper.
A bit more motion: +15%.
Same remix: prompt following, loudness and dialogue unchanged, and dialogue lines still come out word for word. Colour is the same on average, though a few scenes come out less saturated.
Blind A/B: in a blind, AI-judged comparison of full-resolution frames, OmniNFT at the strength v1 bakes in won 29 to 9 against beta 2, with 10 ties.
I tested VBVR too. It cost motion and didn’t improve prompt following, so it isn’t in.
How to run
Put it in
ComfyUI/models/diffusion_modelsand load it with Load Diffusion Model, or pick it in the Models node of Stubelius Ultimate LTX 2.5.CFG 1, 8 steps (plus 3–4 steps if you run an upscale pass). The distilled LoRA and OmniNFT are already baked in, so don’t add either.
Text encoder:
gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot.safetensors. Lightricks re-uploaded it in mid-August, so download it again if your copy is older. VAEs:ltx-2.5-video-vae-bf16andltx-2.5-audio-vae-bf16.The int8 file needs a ComfyUI recent enough to load Comfy-Org’s int8 LTX 2.5 files.
All files: https://huggingface.co/Stuubs/Stubelius_Remix_Ltx2.5
💜 Get new models first on Patreon → https://www.patreon.com/Stuubzzz
🚀 New models and workflows land on Patreon first. Stubelius Remix v1 went to Studio patrons first.
🧩 Workflow Access: my ComfyUI workflows, new ones first. The 18+ tiers add the NSFW workflows.
🎓 Studio: everything above plus the full ComfyUI From Zero course.
💬 Patrons also get the Discord.
🎓 Learn ComfyUI properly: ComfyUI From Zero → https://stuubzzz.studio/comfyui-course/
3 parts, just over 10 hours. Part 1 is free. The full course is €29, paid once, or included in the Studio tier.
📺 Free tutorials and model tests on YouTube → https://www.youtube.com/@StuubzzzCreativeStudios
LTX 2.5, MiniMax H3, LoRA training and more. Subscribe so you catch the next one.
LTX 2.5 Stubelius Remix — beta 2 out now!
Int8 merged with Distilled lora so you can run it on its own + other finetunes.
LTX 2.5 Stubelius Remix — beta 1
A custom LTX 2.5 checkpoint that brings the 2.3-era stylized/NSFW lineage onto the 2.5 engine — built because existing 2.5 ports lean hard into photoreal humans and lose the 2D/stylized soul.
Method (short version): task-vector delta arithmetic, not a naive merge. Deltas extracted from Sulphur 2 (concept & composition) and 10Eros (2D/stylized prior) against their LTX 2.3 base, then injected into the LTX 2.5 transformer with per-block-depth and per-component weighting — early blocks preserve composition knowledge, late blocks favor the stylized prior, photoreal MLP contributions deliberately suppressed. 4,349 tensors, 48 blocks.
Usage
Load as a diffusion model (UNETLoader / Load Diffusion Model) with your normal LTX 2.5 VAEs + text encoder
Run with the LTX distilled LoRA at 0.6–1.0 strength, 10–12 steps. Without it, output is blurry/underbaked (standard 2.5-dev behavior)
CFG 1 with the distill LoRA; your usual samplers (euler / linear_quadratic tested)
Fully compatible with the Stubelius LTX Director pack (seed hunt, refine, Licon MSR references, IC-LoRA, ingredients)
What it's for
2D / 2.5D / anime-styled characters on the full LTX 2.5 feature set (motion, audio, dialogue). NSFW-capable across styles via the Sulphur/Eros lineage. Beta 1: style balance is a first pass and will evolve — feedback steers beta 2.
Credits
TenStrip — Sulphur 2 and 10Eros, the models this stands on
Lightricks — LTX 2.5 base (LTX-Video community license applies)
Merge method & balance — Stubelius
Adult-capable model — intended for consenting-adult creative work only. Check the base and donor model licenses before commercial use.
Description
Beta 2 update After some tweaks following more testing of the Beta 1, I'm finally ready to release the Beta 2 version. As with beta 1, it's mainly meant for NSFW work, but can also do SFW (however you might get some unwanted nudity if not careful with prompts).
This beta 2 has been tuned more towards i2v and it's generally a bit better with anatomy. It also runs better with MSR and Ingredients loras than beta 1.
Run with standard settings: CFG 1, 8steps first pass + 3-4 steps upscale pass. As this model is more tuned towards i2v, general motion quality is improved and coherence should stay intact throughout the generation. Tested on the official LTX2.5 workflow as well as Stubelius ltx2.5 workflow.
For best prompting, you can refer to 10eros prompting guide or Sulphur 2.
Additionally I now merged the Ltx2.5 distilled lora into the model so you don't need to run a separate distilled lora anymore.
FAQ
Comments (22)
Great job. I'm getting quality gens out of this. I'm running a 4070 w/64gb.
Happy to hear that! Before I work on the final version, I will update my workflow to better accommodate the generations so stay tuned for that as well :)
Please don't merge the distill lora into it. Its useless for training loras on top of it this way.
You can extract the lora out of the model and "de-distill it". I wanted to merge the distill lora into it because the model was derived from the DEV model which can be run on fewer steps only with a distill lora. Ofc another option would be to always run a distill lora separately, but for convenience sake it merged it. For lora training, you are right, but you can still train a lora on Ostris AI toolkit for the LTX2.5 base.
Anyways, running a de-distill is super fast (You can ask an AI agent to do it for you)
I have been waiting for this, so far so good, really good, see clips below, no Lora used
Thank you! Let me know if something doesn't work so I can improve for the final version!
@Stuubzzz sure thing, I bet other Loras will work good on very low strength, usually I try with 0.3, in one of my test I tried furry and futanari, how friendly is your checkpoint for either? The furry turned human but it might have been the prompt, the futa test turned the tip later on to mush, but I had the convrot vae and w8 clip, checking it later again, but any recommendations are appreciated 👍
@LynaParker Can you tell me a bit more what mode were you doing? MSR, i2v, t2v, ingredient lora? Helps me pinpoint what the issue could be.
@Stuubzzz hi, no Lora, I2V, plain and simple, not sure what MSR is
@LynaParker Ok thanks for clarification. LTX2.5 is super heavily leaned towards realism and it can twist things. Careful prompting for style will help you and running a euler sampler. Other than that, because of the strong bias towards realism, you often need a lora to combat it. You could try an ltx2.3 furry or *****nari lora with higher strength to see what happens. For me that works quite well. For future I do plan on merging some 2.5d, animation styles into the model, but that takes a long time since training those loras for ltx2.5 isn't exactly fast, sadly.
@Stuubzzz hi, ok, thanks, i will try, with better prompt and furry helper, you asked me if i found something worth mentioning, I think I did, even on higher resolution, there is some sort of flickering or smear or glitch, not on every render, but it happens often, maybe something to smooth out, check out this uploade, 6 clips in total on one edit: https://civitai.red/images/143788643
@LynaParker thanks! That seems to be a latent upscale issue. Will look into it! Thanks!
@LynaParker Update. I have identified the issue. It's from VAE decoding and how the temporal size and overlap create seams at particular frames thus causing the flickering. I'm patching my workflow with the fix and also working on improving the workflow while at it. Thanks again!
@Stuubzzz hi, that is great but I did not use your WF, but let me know what helped, so I can adjust my VAE
Thanks very much for the int8 model! some I2V tests posted ; I couldn't manage to make it work with this workflow which can inject multiple images as keyframes just fine with the basic model+10eros 2.3 lora ; does this model support middle & last frame?
I'm running this model in a director workflow. For example whatdreams cost or my own. In that setting it works normally. So there should not be any model reason why the fml frame workflow shouldn't work...
If you are rocking my Stubelius Director workflow, there is now a patched update that fixes the flickering of the upscale pass. It also now has accurate denoising which reflects the number you actually input as the denoise value (prev. denoise 0.42 was registered as 0.98 due to an error in the calculations).
To get the update, install the workflow if you haven't already, and git pull.
https://github.com/stuubszzz/Stubelius-LTX-Director
very good...waiting for the upgrade
can you (or someone else) host gguf files for this please?
Thanks