protoLabsAI/LTX-2.3-22B-distilled-NVFP4

LTX-2.3-22B-distilled — NVFP4 (mixed precision)

A native NVFP4 (4-bit) quantization of Lightricks/LTX-2.3 22b-distilled-1.1, for NVIDIA Blackwell (RTX PRO 6000 / RTX 50-series, sm120) via ComfyUI.

Lightricks ships an fp4 checkpoint for LTX-2 19B but the newer LTX-2.3 22B is bf16-only. This fills that gap: the 22B distilled model at ~half the disk and VRAM, running on Blackwell's native FP4 TensorCores.

Results (measured, honest)

Single-clip, 960×544, 8-step distilled sampler, distilled decode, RTX PRO 6000 (96 GB), vLLM-adjacent ComfyUI stack on torch cu130. Speed/VRAM reproduced across 2 runs each; quality eyeballed across 4 prompts.

Metric	bf16 (upstream)	NVFP4 (this)	Δ
Disk size	46.1 GB	22.9 GB	2.0× smaller
DiT step speed	2.85 s/it	1.82 s/it	1.57× faster
Peak runtime VRAM	~60 GB	~37 GB	−38%
Cold load + first clip	39.2 s	19.0 s	2.1× faster
Distilled-decode quality	baseline	visually unchanged	—

Higher resolution holds the win (fp4 1280×704 ≈ 3.85 s/it, scaling with token count as expected).

⚠️ You need torch built with CUDA 13 (cu130)

This is the single most important thing. NVFP4 only hits the fast TensorCore path on cu130. On cu128, fp4 silently falls back and runs ~2× slower than fp8 (ComfyUI logs WARNING: You need pytorch with cu130 or higher to use optimized CUDA operations). This is the field's #1 LTX-fp4 confusion. If your fp4 is slower than fp8, this is why.

pip install --force-reinstall --no-cache-dir torch==2.11.0 torchvision torchaudio \
  --index-url https://download.pytorch.org/whl/cu130

⚠️ Use the distilled decode path

The LTX-2.3 single-stage workflow ships two decode branches (a fast "distilled" decode and a "full" decode). On this quant, the distilled decode is visually indistinguishable from bf16, while the full decode shows mild added artifacting. Default to the distilled decode. (The DiT is where the fp4 lives; the VAE decoders are untouched bf16 — the full decoder is just more sensitive to the small DiT-linear error.)

Usage (ComfyUI)

1. cu130 torch (above) + ComfyUI-LTXVideo (provides the comfy_kitchen NVFP4 kernels: scaled_mm_nvfp4, quantize_nvfp4).
2. Drop ltx-2.3-22b-distilled-1.1-fp4.safetensors into ComfyUI/models/checkpoints/.
3. Load the 2.3/LTX-2.3_T2V_I2V_Single_Stage_Distilled_Full example workflow, point the CheckpointLoaderSimple at this file, keep the distilled decode/save branch.
4. Text encoder: any Gemma-3-12B LTX encoder (gemma_3_12B_it_fp8_scaled.safetensors works well).

The checkpoint carries a standard _quantization_metadata header, so ComfyUI's LTX loader auto-detects the NVFP4 layers — no flags, no config edits.

What's quantized (mixed precision)

Replicates Lightricks' own 19B-fp4 layer policy exactly (reverse-engineered from their shipped checkpoint):

• NVFP4 (4-bit): the 2-D Linear weights in transformer_blocks 1–42 (attention q/k/v/out + feed-forward, video + audio + cross-modal). 1,176 weights.
• Kept bf16 (precision-sensitive): transformer block 0 and the last 5 blocks (43–47), all to_gate_logits gates, the patchify / proj_out / adaln / caption + embeddings connectors, and the entire VAE / audio-VAE / vocoder.

Each fp4 weight is stored as W (uint8 packed) + W_scale (fp8-e4m3 per-block) + W_scale_2 (fp32 per-tensor) — byte-identical to the shipped 19B fp4 format.

Reproduce it

The full converter is quantize.py (in this repo / protoLabsAI/lab). It streams the bf16 checkpoint tensor-by-tensor through comfy_kitchen's own TensorCoreNVFP4Layout.quantize, so the output format is guaranteed loader-compatible. ~40 s on one Blackwell card.

CUDA_VISIBLE_DEVICES=0 python quantize.py \
  --in  ltx-2.3-22b-distilled-1.1.safetensors \
  --out ltx-2.3-22b-distilled-1.1-fp4.safetensors

Limitations & honest notes

• Numbers are single-clip on one specific Blackwell + cu130 ComfyUI config; your mileage varies with resolution, length, sampler, and card.
• Speedup is on the DiT denoising loop. End-to-end wall-clock also includes VAE decode (bf16, unchanged) and text encoding, so the whole-pipeline ratio is smaller than 1.57× at short clip lengths.
• Full-decode path: mild artifacting (see above). Distilled decode only.
• FP4 is 4-bit — expect a small per-tensor error (~9% relative L1 on a quantized layer), which the first/last bf16 blocks are there to absorb. If you see quality loss, widen the bf16-kept block set and re-quantize.

Attribution & license

Derivative of Lightricks/LTX-2.3, released under the LTX-2 Community License (see LICENSE). This quant is redistributed under the same license, including its Acceptable Use Policy (Attachment A) and paragraph 4 use restrictions, which pass through to you. High-revenue commercial entities require a separate Commercial Use Agreement from Lightricks for use — see the license.

Quantized by protoLabsAI. Method + measurements: protoLabsAI/lab → experiments/ltx2-nvfp4/.