ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors
LTX 2.3 Dev FP8 Scaled (Kijai)
FP8 scaled dev model by Kijai. 16GB VRAM. Supports LoRA. Place in models/checkpoints/.
Released 2026-03 · Source: Kijai/LTX2.3_comfy (HuggingFace) — Part of Kijai's initial LTX 2.3 ComfyUI port. The FP8 scaled and fp8_input_scaled variants shipped together as alternative quantization paths.
Download ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors
Direct HuggingFace download. ~25 GB · Free.
No 16GB GPU? Try ltx-2.3-22b-dev_transformer_only_fp8_scaled.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
FP8 scaled quantization stores transformer weights as 8-bit floats with per-channel scale tables. Compared to fp8_input_scaled, this variant applies scaling on the weight side rather than the input side — the result is slightly different numerical behavior in attention, with no consistent quality winner; both are very close to BF16 reference.
This is the dev (non-distilled) base model at FP8 size. That means 30-step sampling with CFG > 1 is the default, LoRA application works cleanly, and you have flexibility the distilled checkpoints don't give you. The file is ~25 GB and fits in 16 GB VRAM with the FP4 Gemma text encoder.
Kijai's transformer_only files like this one are designed for the standard ComfyUI workflow pattern: load the transformer, VAE, and text encoder as separate nodes. Lightricks' first-party FP8 files (Lightricks/LTX-2.3-fp8) are full checkpoints with everything bundled — they're a different shape entirely.
When to choose ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors
Use this when you want the dev model at FP8 size with hardware-accelerated matmul on RTX 40xx+ (Ada or Blackwell), and the standard fp8_input_scaled variant has caused issues for your workflow — typically when a LoRA loader complained about input scaling, or when output looked subtly off in I2V conditioning.
For most users, ltx-2.3-22b-distilled-1.1_transformer_only_fp8_scaled.safetensors is the better default — it's the distilled v1.1 (newer weights, 8-step inference, 4× faster). Pick this dev FP8 only when you need LoRA flexibility that the distilled path doesn't give cleanly, or when training-related work demands the dev base.
On RTX 30-series, switch to ltx-2.3-22b-dev_transformer_only_mxfp8_block32.safetensors — same role, format that runs on Ampere.
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: Alternative FP8 quantization for dev. Use when fp8_input_scaled has compatibility issues.
How to use ltx-2.3-22b-dev_transformer_only_fp8_scaled.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.
Don't want to run this locally? Try ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors online with a free generation — no GPU, no install, ~30 seconds per clip.
ComfyUI says it can't find ltx-2.3-22b-dev_transformer_only_fp8_scaled.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-dev_transformer_only_fp8_scaled.safetensors
- ltx-2\ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors
- diffusion_models/ltx-2.3-22b-dev_transformer_only_fp8_scaled.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-dev_transformer_only_fp8_scaled.safetensors — you have two fixes:
- Create the matching subdirectory inside ComfyUI/models/checkpoints/ and place the file there. Example: if the workflow references ltx\ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors, create the corresponding subfolder under ComfyUI/models/checkpoints/ and put ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors inside it.
- Or open the workflow JSON in a text editor and replace the prefixed string with just ltx-2.3-22b-dev_transformer_only_fp8_scaled.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
Output is subtly different from the same prompt with fp8_input_scaled▼
Different scaling axis. fp8_scaled scales weights; fp8_input_scaled scales activations on the way in. Both are valid quantizations but produce slightly different attention numerics. Fix: Pick one and stick with it. Don't expect to swap between them mid-pipeline and get identical output. If you need exact reproducibility, the BF16 dev file is the only truly stable reference.
RuntimeError: fp8 matmul not supported on RTX 3090▼
Ampere lacks native FP8 matmul tensor cores. Fix: Switch to ltx-2.3-22b-dev_transformer_only_mxfp8_block32.safetensors. Same VRAM, runs on RTX 30-series.
LoRA application produces visual artifacts (banding, color shift)▼
Some LoRAs were trained against fp8_input_scaled and apply their deltas in a way that doesn't compose cleanly with weight-scaled FP8. Fix: Try the LoRA against the BF16 dev model first to confirm it works at all. If it's specifically tied to input-scaled, switch to the fp8_input_scaled variant. Or apply LoRA against BF16 and quantize after, if your tooling supports it.
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-dev_transformer_only_fp8_scaled.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-dev_transformer_only_fp8_scaled.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 Dev FP8 Scaled (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-1.1_transformer_only_mxfp8_block32.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_mxfp8_block32.safetensors
ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled.safetensors