Stuubs/Stubelius_Remix_Ltx2.5

LTX 2.5 Stubelius Remix

v1 update (2026-10-07)

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 changed, 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; 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.

Files

File	Size
ltx2.5-Stubelius_remix_v1_int8_convrot.safetensors	21.5 GB	v1, the one to use (ComfyUI int8)
ltx2.5-Stubelius_remix_v1_bf16.safetensors	42 GB	v1 in bf16, for merging and your own conversions
ltx2.5-Stubelius_remix_v1_nvfp4.safetensors	14.4 GB	v1 in NVFP4 for RTX 50-series (Blackwell) cards, experimental (see below)
ltx2.5-Stubelius_remix_beta2_int8_convrot.safetensors, ltx2.5-Stubelius_remix_beta2_bf16.safetensors	21.5 / 42 GB	beta 2
ltx2.5-Stubelius_remix_beta1.safetensors	42 GB	beta 1

How to run v1

• Put it in ComfyUI/models/diffusion_models and 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; download it again if your copy is older. VAEs: ltx-2.5-video-vae-bf16 and ltx-2.5-audio-vae-bf16.
• The int8 file needs a ComfyUI recent enough to load Comfy-Org's int8 LTX 2.5 files. LTX 2.5 LoRAs load on it the same way as on stock LTX 2.5.

NVFP4 (experimental)

A smaller 4-bit version of the same v1 weights for RTX 50-series cards: transformer blocks 0–41 in NVFP4, the last 6 blocks and the text connectors in int8. That's Lightricks' own layer choice for their NVFP4 file, with int8 where they kept bf16.

• Same settings as int8: CFG 1, 8 steps. Load it the same way.
• It's only fast on RTX 50-series (Blackwell) cards, with a recent ComfyUI (0.32 or newer) and PyTorch built for CUDA 13. Older cards can run it, but slower than the int8 file, so stay on int8 there.
• It's new. The weights check out: every 4-bit layer is within normal NVFP4 rounding, and the OmniNFT change is kept. I haven't finished comparing its renders against int8 yet, though, so if something looks off, switch to int8 and tell me.

Credits and licence

• Lightricks: LTX 2.5 and the distilled LoRA
• FusionCow / SulphurAI: Sulphur 2
• TenStrip: 10Eros
• zghhui et al.: OmniNFT (Apache-2.0); Kijai: its ComfyUI conversion
• Merge method and balance: Stuubzzz

These models are derivatives of LTX-2.5 under the LTX-2.x Community License Agreement. By downloading or using them you agree to it, including its use restrictions (Section 4 and Attachment A). Companies with USD 10M+ annual revenue need a licence from Lightricks first. NOTICE.txt lists what was changed and the other notices.

Adult-capable model — intended for consenting-adult creative work only. Check the base and donor model licenses before commercial use.

Beta 2 update (2026-09-19)

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.

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.