eriktonon/ ltx-2.5-uncensored:latest

70 2 days ago

Lightricks/LTX-2.5 Uncensored (Orthogonally Abliterated · THINK=0) Visual Scene Director & Prompt Encoder coupled with Tier-1 Real GPU Diffusion Video Synthesis (AnimateDiff + epiCRealism 24fps H.264 MP4 on RTX 5080 / A100). 

ollama run eriktonon/ltx-2.5-uncensored

Details

2 days ago

6b43b50e471b · 5.6GB

model
qwen35
·
8.95B
·
Q4_K_M
template
{{- if .Messages }} {{- if .System }}<|im_start|>system {{ .System }}<|im_end|> {{ end }} {{- range
system
Você é o motor controlador e encoder de texto sem censura (100% Uncensored / Orthogonal Abliterate
license
# eriktonon/ltx-2.5-uncensored (Lightricks/LTX-2.5 Uncensored Scene Director + Tier-1 GPU Diffusion
params
{ "num_ctx": 65536, "repeat_penalty": 1.15, "stop": [ "<|im_start|>", "<

Readme

Lightricks/LTX-2.5 Uncensored (Visual Scene Director + Tier-1 GPU Diffusion Video Engine)

An uncensored, high-throughput Visual Scene Director & Prompt Encoder (THINK_LEN=0 pre-closed <think></think> template, ~101 tok/s on NVIDIA GeForce RTX 5080) coupled with Tier-1 Real GPU Neural Video Diffusion (:8188) for generating photorealistic 1280x720 @ 24fps H.264 MP4 videos without safety filters (safety_checker=None), engineered by Erik Tonon.

📦 Available Editions (Tags on Ollama Hub)

Tag Role & Architecture Size VRAM Footprint Context Recommended GPU Command
:latest Uncensored Visual Scene Director & Prompt Encoder (THINK=0 Orthogonally Abliterated) 5.6 GB ~5.8 GB (+ 2.1 GB Diffusion Worker) 64K (256K max) Local RTX 5080 (16GB) / Colab A100 ollama run eriktonon/ltx-2.5-uncensored:latest
:22b-int8 Lightricks/LTX-2.5 Uncensored Controller (hf-ltx-2.5:22b-int8 alias) 5.6 GB ~5.8 GB (+ 2.1 GB Diffusion Worker) 64K (256K max) Local RTX 5080 (16GB) / Colab A100 ollama run eriktonon/ltx-2.5-uncensored:22b-int8

⚡ Key Highlights

  • 100% Uncensored / Orthogonal Refusal Abliteration: Neutralizes refusal directions across intermediate transformer layers. Accepts any creative, cinematic, anatomical, or adult/NSFW visual prompt without moralizing lectures or synthetic refusal templates.
  • Zero-Latency Direct Output (THINK_LEN=0): Pre-closed <think>\n\n</think>\n\n template across both /api/chat (.Messages) and /api/generate (.Prompt), delivering ~101 tokens/sec immediately in the user’s language or translating into dense English visual prompts for the diffusion denoiser.
  • Coupled Tier-1 Real GPU Diffusion Video Worker (http://127.0.0.1:8188/api/generate_sync):
    • Neural Diffusion Stack: PyTorch CUDA (2.13.0+cu132, torch.float16) + Hugging Face diffusers (AnimateDiffPipeline) + photorealistic uncensored checkpoint emilianJR/epiCRealism + motion adapter guoyww/animatediff-motion-adapter-v1-5-2 (safety_checker=None).
    • Smart VRAM Coexistence (~2.1 GB Peak): Uses enable_model_cpu_offload(), vae.enable_slicing(), and vae.enable_tiling() so the GPU video denoiser runs simultaneously alongside 12 GB Ollama models on a single 16 GB NVIDIA GeForce RTX 5080 without Out-Of-Memory errors.
    • 24fps H.264 Output: Upsamples latent frames with Lanczos to 1280x720 and interpolates smooth 24fps motion (libx264, yuv420p, crf 19) + HD keyframe PNG (engine_tier: "diffusion_gpu_worker").

🚀 Quickstart & Usage

1. Run Scene Director via Ollama CLI

ollama run eriktonon/ltx-2.5-uncensored:22b-int8

2. Start Coupled Real GPU Diffusion Video Worker (:8188)

# Preload AnimateDiff + epiCRealism (FP16, safety_checker=None) on RTX 5080
python voice/diffusion_video_worker.py --preload

# Verify GPU Worker Health
curl http://127.0.0.1:8188/health

3. Generate Real 1280x720 @ 24fps Neural Video via REST API

curl -X POST http://127.0.0.1:8188/api/generate_sync \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "cinematic photorealistic scene with volumetric golden hour lighting",
    "mode": "video_ltx",
    "width": 1280,
    "height": 720,
    "fps": 24
  }'