ltx-2.3-22b-distilled-1.1_transformer_only_mxfp8_block32.safetensors
LTX 2.3 Distilled 1.1 MXFP8 (Kijai)
MXFP8 block-32 quantized distilled 1.1 by Kijai. Use on RTX 30xx GPUs that cannot run standard FP8 scaled matmul.
Released 2026-04-13 · Source: Kijai/LTX2.3_comfy (HuggingFace) — v1.1 release. The v1.1 line improved fast-motion stability over v1.0 and dropped on the same date as the v1.1 FP8 scaled file.
Download ltx-2.3-22b-distilled-1.1_transformer_only_mxfp8_block32.safetensors
Direct HuggingFace download. ~25 GB · Free.
No 16GB GPU? Try ltx-2.3-22b-distilled-1.1_transformer_only_mxfp8_block32.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.
Technical details
MXFP8 (Microscaling FP8, block-32) is a quantization format that stores 32-value blocks each sharing a single FP8 scale factor. The key property is that the matmul kernels work on standard BF16 tensor cores — you don't need the Ada / Hopper / Blackwell-specific FP8 matmul instructions that ltx-2.3-22b-distilled-1.1_transformer_only_fp8_scaled.safetensors requires.
File size matches the FP8 scaled variant at roughly 25 GB. Quality is also very close — the block-32 grouping preserves enough dynamic range that distilled output is visually indistinguishable from the FP8 scaled file in side-by-side comparisons on standard prompts.
'transformer_only' means the file contains only the DiT weights. Pair it with taeltx2_3.safetensors (VAE) and a Gemma 3 12B text encoder (FP4 mixed on 16 GB, FP8 scaled or BF16 on more). All current ComfyUI workflows reference this filename verbatim from Kijai's repo.
When to choose ltx-2.3-22b-distilled-1.1_transformer_only_mxfp8_block32.safetensors
Pick MXFP8 specifically if your GPU is RTX 30-series (3060 12GB, 3080 10/12GB, 3090 24GB) or older Ampere/Turing. Those cards lack native FP8 matmul tensor cores — running the fp8_scaled variant either crashes with an unsupported-dtype error or falls back to a slow emulation path that negates the speedup.
On RTX 40-series (Ada) or RTX 50-series (Blackwell), use the standard FP8 scaled variant instead — same VRAM, same quality, slightly faster because the matmul uses dedicated FP8 hardware.
This is the v1.1 release. The v1.0 MXFP8 (without -1.1 in the name) is older and worse for fast camera motion — switch up unless you're reproducing a specific v1.0 result.
Will this run on my GPU?
Minimum: 16GB VRAM. Headroom up to: 24GB.
⚠ 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: RTX 30xx workaround — use this when your GPU lacks RTX 40xx-style FP8 matmul support. Same VRAM as fp8_scaled.
How to use ltx-2.3-22b-distilled-1.1_transformer_only_mxfp8_block32.safetensors
- Download the file from HuggingFace.
- Place it in ComfyUI/models/checkpoints/ inside your ComfyUI directory.
- Restart ComfyUI (or refresh the model list from the menu).
- Load a compatible workflow — see below.
Compatible official workflows:
- LTX-2.3_T2V_I2V_Single_Stage_Distilled_Full.json— T2V / I2V Single Stage Distilled
- LTX-2.3_T2V_I2V_Two_Stage_Distilled.json— T2V / I2V Two Stage Distilled
- LTX-2.3_ICLoRA_Union_Control_Distilled.json— ICLoRA Union Control Distilled
- LTX-2.3_ICLoRA_Motion_Track_Distilled.json— ICLoRA Motion Track Distilled
- LTX-2.3_ICLoRA_HDR_Distilled.json— ICLoRA HDR Distilled
Don't want to run this locally? Try ltx-2.3-22b-distilled-1.1_transformer_only_mxfp8_block32.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-1.1_transformer_only_mxfp8_block32.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:
- ltx23\ltx-2.3-22b-distilled-1.1_transformer_only_mxfp8_block32.safetensors
- diffusion_models/ltx-2.3-22b-distilled-1.1_transformer_only_mxfp8_block32.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-1.1_transformer_only_mxfp8_block32.safetensors — you have two fixes:
- Create the matching subdirectory inside ComfyUI/models/checkpoints/ and place the file there. Example: if the workflow references ltx23\ltx-2.3-22b-distilled-1.1_transformer_only_mxfp8_block32.safetensors, create the corresponding subfolder under ComfyUI/models/checkpoints/ and put ltx-2.3-22b-distilled-1.1_transformer_only_mxfp8_block32.safetensors inside it.
- Or open the workflow JSON in a text editor and replace the prefixed string with just ltx-2.3-22b-distilled-1.1_transformer_only_mxfp8_block32.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
Runs but is no faster than BF16 on my RTX 3090▼
MXFP8 dequantizes to BF16 for the matmul itself — the speedup comes from halved memory bandwidth, not faster compute. On a 24 GB card with the full pipeline in memory, you're already mostly compute-bound, not memory-bound. Fix: This is expected. The win on RTX 30-series is fitting the model in VRAM at all — a BF16 transformer would be ~44 GB and OOM on 24 GB without sequential offloading. Use the FP8 scaled file on RTX 40xx+ for actual compute speedup.
ComfyUI 'Mismatched shapes' error when stacking with a LoRA▼
Some older LoRA loaders don't understand MXFP8 weight layout and try to apply LoRA deltas in the wrong dtype. Fix: Update ComfyUI to a recent version (post-2026-04) and ensure you're using KJNodes if your workflow needs Kijai-specific loaders. Or apply the LoRA against the BF16 transformer instead and quantize after, if your trainer supports it.
Black or noisy first frame, rest of video looks fine▼
Workflow loaded the MXFP8 file with a node configured for fp8_scaled or BF16 — internal scale tables aren't being applied to the first denoising step. Fix: Use ComfyUI's standard CheckpointLoaderSimple or the Kijai LTXVideoModelLoader from KJNodes. Avoid custom loaders that hard-assume a specific dtype.
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-1.1_transformer_only_mxfp8_block32.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-1.1_transformer_only_mxfp8_block32.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.
Get notified when LTX 2.3 Distilled 1.1 MXFP8 (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
ltx-2.3-22b-distilled-1.1_transformer_only_fp8_scaled.safetensors
ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors
ltx-2.3-22b-dev_transformer_only_fp8_input_scaled.safetensors
ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors
ltx-2.3-22b-dev_transformer_only_mxfp8_block32.safetensors
ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled.safetensors