ChrisColeTech/LTX-2.3-uncensored-fp8

LTX-2.3 22B uncensored — GGUF / FP8 / INT8

The 22B LTX-2.3 transformer in six quantizations, plus the text encoders, VAEs, distill LoRAs and spatial upscaler it needs. Weights only — not a retrain.

What this repo is: a complete split layout for running LTX-2.3 locally. The settings below are the values these weights are run with day to day, measured on this hardware, not copied from the upstream card.

---

Sample

_{prompt: a red apple on a rustic wooden table beside a window, soft daylight, the camera slowly pushes in, dust motes drifting through the light, cinematic, shallow depth of field — 1024×576, 121 frames, 24 fps, 12 steps, guidance 1.0, seed 42, negative prompt as below. 5.04 s of video generated with the Q4_K_M transformer in ≈11 min on an RTX 5090 (32 GB), including model load.}

_{Frames sampled at 0.2 s / 1.6 s / 3.2 s / 4.8 s. The apple grows across them — that is the prompted camera push-in, not a still. Dust motes drift through the window light, and the depth-of-field falloff holds across the whole clip.}

---

Recommended settings

Parameter	Value	Notes
width × height	1024 × 576	see the resolution note below
num_frames	121	5.04 s at 24 fps
frame_rate	24.0
steps	12
guidance	1.0	distilled — CFG is not used
negative_prompt	worst quality, blurry, jittery, distorted, static

Supported modes: txt2video, img2video

⚠ Use 1024×576, not 768×512

The upstream card suggests 768×512. At 121 frames that resolution reproducibly desaturates: measured colour spread of roughly 7 across four seeds, against ≈29 at 1024×576 and at every other resolution tested.

If output looks washed out, check the resolution before anything else.

Steps: 12, not 8

Measured sharpness (mean Laplacian variance) at 1024×576×121 rises 66.5 → 125.3 going from 8 to 12 steps, with equal colour. 12 is the preset.

---

Files

File	Size	Role
split/unet/ltx-2.3-22b-uncensored-Q4_K_M.gguf	14.30 GB	transformer — the quant the sample above was generated with; fits a 32 GB card comfortably
split/unet/ltx-2.3-22b-uncensored-Q6_K.gguf	17.77 GB	transformer, higher-precision quant
split/unet/ltx-2.3-22b-uncensored-Q8_0.gguf	22.76 GB	transformer, highest-precision quant
split/unet/ltx-2.3-22b-uncensored-fp8.safetensors	29.16 GB	transformer, scaled FP8
split/unet/ltx-2.3-22b-uncensored-int8.safetensors	29.16 GB	transformer, INT8
split/unet/ltx-2.3-22b-distilled-fp8.safetensors	29.53 GB	the distilled (non-uncensored) transformer, FP8
split/text_encoders/gemma-3-12b-it-qat-UD-Q4_K_XL.gguf	7.43 GB	Gemma-3-12B text encoder
split/text_encoders/ltx-2.3-22b-dev_embeddings_connectors.safetensors	2.31 GB	embedding connectors
split/text_encoders/mmproj-BF16.gguf	0.85 GB	multimodal projector
split/lora/ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors	0.66 GB	distill LoRA — required, applied at load
split/lora/ltx-2.3-22b-distilled-lora-384-1.1.safetensors	7.61 GB	distill LoRA, 384-rank variant
split/vae/ltx-2.3-22b-distilled_video_vae.safetensors	1.45 GB	video VAE
split/vae/ltx-2.3-22b-distilled_audio_vae.safetensors	0.36 GB	audio VAE
split/latent_upscale_models/ltx-2.3-spatial-upscaler-x2-1.0.safetensors	1.00 GB	2× spatial upscaler
split/tokenizer/	38 MB	Gemma tokenizer

A single working set is about 27 GB — one transformer (Q4_K_M) + Gemma + connectors + both VAEs + one distill LoRA + tokenizer. Downloading every transformer variant is ≈144 GB and is not necessary.

---

Notes and gotchas

• A distill LoRA is required, not optional. The pipeline resolves it as a mandatory component and will not load without one. Keep it in a split/distill_loras/ directory if your loader discovers components by directory role — a split/lora/ name will leave the slot unfilled and the model will be silently skipped.
• The audio VAE is present but the sample has no audio. Audio generation is a separate path and is not exercised by the settings above.
• guidance stays at 1.0. This is a distilled model; it is not trained for classifier-free guidance, and raising the value is not the way to increase prompt adherence.
• Generation is slow relative to image models — ≈11 minutes for 5 s of video at 12 steps and Q4_K_M on an RTX 5090, most of it in the diffusion loop rather than model load. Expect roughly linear scaling with num_frames.
• Note Q6_K (17.77 GB) and the distilled fp8 (29.53 GB) are different base checkpoints, not merely different quants of one file — the distilled-fp8 entry is the 1.0 distill, the uncensored-* entries are the 1.1 line.

---

Provenance

• Upstream base model: LTX-2.3 22B (Lightricks)
• This build: quantizations of the uncensored and distilled checkpoints, redistributed with the text encoders, VAEs, distill LoRAs and upscaler alongside. Weights are not retrained here.
• License: left as unknown in this repo's metadata. Refer to the upstream model's license for redistribution and commercial-use terms.