LTX 2.3 22B — Pure noise/garbage output in ComfyUI

[EGBluesman]

Running the LTX 2.3 22B transformer-only model in ComfyUI produces completely garbled output (pure noise, no recognizable content) regardless of prompt.

Setup:

• Model: ltx-2.3-22b-dev (transformer only)
• VAE: Video VAE + Audio VAE loaded separately
• Text encoder: gemma_3_12B_it_fp8_e4m3fn.safetensors (loaded via GGUF CLIP loader node)
• Text projection: ltx-2.3_text_projection_bf16.safetensors
• ComfyUI version: [0.19.1]
• GPU / VRAM: [4060TI 16GB]
• OS: Windows 11

What I expected

A generated video matching the prompt

What I got

Completely garbled frames — uniform noise across the entire output, every time

What I have tried

• Multiple different workflows from YouTube tutorials
• Switching from fp4_mixed to fp8_e4m3fn Gemma variant (resolved a shape mismatch error but output still garbage)
• Multiple different prompts

Appreciate someone can help
Thank you all

[FilipeF12]

This is what I get. You getting the same?

Model: ltx-2.3-22b-dev (distilled transformer)
VAE: Video VAE + Audio VAE loaded separately
Text encoder: gemma_3_12B_it_fp8_scaled.safetensors
Text projection: ltx-2.3_text_projection_bf16.safetensors
LoRA: ltx23_anime2real_rank64_v1_4500 (1.0)
ComfyUI version: [0.19.1]
GPU / VRAM: [3090 24GB]
OS: Windows 10

WF: https://huggingface.co/Alissonerdx/LTX-LoRAs/tree/main/workflows/ltx23_anime2real_v1.json

[lycos2019]

same problem here

[FilipeF12]

wf says report problems here. like, why.

[QZ-Bloodwolf]

You are using the dev model, if you are only running the default 8 step sigmas then that is the result you will get. You have to load the Distilled lora or up the steps substantially

[FilipeF12]

QZ-Bloodwolf my dude, using the wf you supplied, which now is 404. I kept a copy,

My image above is not the dev model, it's option 2: ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors

To avoid any further confusion, do you have a working wf for us? Thx.

[QZ-Bloodwolf]

FilipeF12 haha you're correct, I checked the OP and was applying his issue with your image, didn't see it was 2 different usernames😂
And I'm not sure what wf you are referencing because I have not provided one?
But the reault would still be similar by running the Dev (non distilled) in 8 steps without the Distilled lora.
So that would make me wonder if in your case the model you are using is actually the distilled one or if there has been a mixup, easy tontest though by slapping on the lora an running a quick test, and making sure to use the Euler_ancestral_cfg_pp sampler as the first pass(can skip the upscale pass when just running the test)

[FilipeF12]

QZ-Bloodwolf haha ok to clear things up, the wf I am using (linked above in my 1st post, now 404) is from hf dude Allisonerdx. The wf states to report wf errors here. I just chained on to EGBluesman's report.

Further, it 404's because either hf or Allisonerdx changed the path. Here's the new correct url: https://huggingface.co/Alissonerdx/LTX-LoRAs/blob/main/workflows/ltx23_anime2real_v1.json

"To report issues when running this workflow, go here: https://github.com/Lightricks/ComfyUI-LTXVideo/issues"
Re-reading that, I think that guy needs to be more clear about what issues to report here.

Not totally sure what went wrong on the first run above, but it's all good now. I made sure it's the distilled version and the correct sampler. I have a better test anime as well. Thanks for your help bro!