25 GB16GB+ VRAMfp8

ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors

LTX 2.3 Distilled FP8 v3 (Kijai)

FP8 distilled v3 by Kijai. Previous version, superseded by v1.1 FP8.

Released 2026-03 · Source: Kijai/LTX2.3_comfy (HuggingFace)Final v1.0-era distilled FP8 quantization. Superseded by the v1.1 line on 2026-04-13.

Download ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors

Direct HuggingFace download. 25 GB · Free.

Install path: ComfyUI/models/checkpoints/ + ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors

No 16GB GPU? Try ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors online — free generation included

Skip the 25 GB download and ComfyUI setup. Generate a 6-second video using this exact model in your browser, ~30 seconds.

Try this model online — free →

Technical details

This is the third iteration of Kijai's FP8 input-scaled distilled quantization, before the v1.1 line existed. The 'v3' suffix tracks Kijai's own re-quantization passes — v1 and v2 had calibration issues that v3 fixed, making this the last and most refined v1.0-era distilled FP8.

'fp8_input_scaled' applies the FP8 scale tables on the activation side rather than the weight side. In practice this preserves a bit more precision in the attention layers when prompts have unusual token distributions, at the cost of slightly more compute overhead at inference time.

The weights underneath are the v1.0 distilled checkpoint. That means 8-step sampling, CFG=1, fast inference — but without the fast-motion and consistency improvements that came in v1.1.

When to choose ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors

Worth keeping only if you're reproducing a specific result from a workflow JSON that hardcodes this exact filename, or doing a v1.0 → v1.1 quality comparison.

For everything else, switch to ltx-2.3-22b-distilled-1.1_transformer_only_fp8_scaled.safetensors. v1.1 is meaningfully better for fast camera motion, character consistency across frames, and prompt adherence. Same VRAM footprint, same inference speed.

If you specifically need the input-scaled style (instead of weight-scaled), there's currently no v1.1 input_scaled release — you'd stay on this file. But most workflows are equally happy with the v1.1 fp8_scaled variant.

Will this run on my GPU?

Minimum: 16GB VRAM. Headroom up to: 24GB.

GPUVRAMVerdict
RTX 3060 12GB12GBInsufficient VRAM
RTX 4060 Ti / 4070 (16GB)16GBTight fit
RTX 4070 Ti SUPER / 4080 (16GB)16GBTight fit
RTX 3090 (24GB)24GBNo FP8 support
RTX 4090 (24GB)24GBComfortable
RTX 5090 / A6000 (32GB+)32GBComfortable

⚠ FP8 scaled matmul requires RTX 40-series or newer (Ada Lovelace architecture). RTX 30xx cannot run this format — use the MXFP8 block-32 or BF16 variant instead.

Recommendation: Previous version. Use v1.1 FP8 for latest quality.

How to use ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors

  1. Download the file from HuggingFace.
  2. Place it in ComfyUI/models/checkpoints/ inside your ComfyUI directory.
  3. Restart ComfyUI (or refresh the model list from the menu).
  4. Load a compatible workflow — see below.

Compatible official workflows:

Don't want to run this locally? Try ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors online with a free generation — no GPU, no install, ~30 seconds per clip.

ComfyUI says it can't find ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors?

Some published workflow JSONs reference this file under a custom subdirectory. If ComfyUI shows a "cannot find model" error and your workflow references one of these path-prefixed variants:

  • ltx\ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors
  • diffusion_models/ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors

The prefix before the slash or backslash is a subdirectory the workflow author used. The actual file is the same ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors — you have two fixes:

  1. Create the matching subdirectory inside ComfyUI/models/checkpoints/ and place the file there. Example: if the workflow references ltx\ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors, create the corresponding subfolder under ComfyUI/models/checkpoints/ and put ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors inside it.
  2. Or open the workflow JSON in a text editor and replace the prefixed string with just ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors. ComfyUI then resolves it directly from ComfyUI/models/checkpoints/.

On Windows the separator is \, on macOS/Linux it is / — they refer to the same nested folder regardless of platform.

Common issues

Workflow JSON references the file without the _v3 suffix

Older workflows targeted v1 or v2 — the filename string is hardcoded and not interchangeable across versions. Fix: Either rename the file to match the workflow's expected string (you'll lose the version tag) or edit the workflow JSON to use the _v3 filename. The latter is cleaner.

Output quality is lower than the v1.1 fp8_scaled file on the same prompt

v3 is v1.0 weights. v1.1 trained on more data with better calibration. Fix: Switch to ltx-2.3-22b-distilled-1.1_transformer_only_fp8_scaled.safetensors. The quality delta is most visible in fast pan/zoom shots and character close-ups.

RuntimeError: fp8 matmul not supported (RTX 30-series)

Ampere lacks FP8 tensor cores. Input scaling doesn't change this hardware requirement. Fix: There's no v1.0 MXFP8 distilled file that matches this exactly. Switch up to ltx-2.3-22b-distilled-1.1_transformer_only_mxfp8_block32.safetensors — v1.1 is better anyway.

ComfyUI doesn't see the file after I downloaded it

Make sure the file is in ComfyUI/models/checkpoints/ (not a subfolder). Restart ComfyUI fully — the menu refresh sometimes misses new files. Filename must match exactly: ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors.

I get a CUDA error mentioning fp8 / scaled / matmul

FP8 scaled matmuls require an RTX 40-series GPU or newer (Ada Lovelace architecture). RTX 30-series and older cannot run FP8 weights at native precision. Use the BF16 variant instead, or the MXFP8 block-32 alternative.

CUDA out of memory error when loading the model

ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors needs ~16GB VRAM minimum. If you're hitting OOM: • Enable Sequential Offloading in ComfyUI settings • Lower the resolution (768×512 instead of 1280×704) — both dimensions must be divisible by 32 • Reduce frame count (65 frames instead of 161) — must be 8n+1 • Use a smaller variant — see Related models below.

Free newsletter

Get notified when LTX 2.3 Distilled FP8 v3 (Kijai) updates

Occasional updates on what's new in LTX 2.3 — new FP8 quants, LoRAs, IC-LoRA releases — with our hands-on verdict on whether they're worth re-downloading. No fixed cadence.

No spam. Sent occasionally when there's real news. Unsubscribe in one click.

Related models