🚀 MrXin LTX 2.3 I2V EROS V6
This workflow has continuously evolved based on user feedback to deliver one of the strongest LTX 2.3 I2V experiences available. Thx everyone for the feedbacks.
V6 is a major leap forward, focusing on stronger motion preservation, better prompt adherence, and more user control while keeping excellent stability and VRAM efficiency.
🔥 What’s New in V6 (Compared to V5)
Added LTX2 NAG (Negative Attention Guidance) on both First Pass and Final Pass for significantly improved prompt following and motion quality.
Switched to VAELoader KJ nodes for Video VAE and Audio VAE (better compatibility and stability).
Improved Final Pass by repositioning the LTXVCropGuides node for cleaner latent processing.
Added flexible sigma system: Manual Sigmas (8 steps) or automatically generated sigmas via LTXVScheduler + Sigmas Sigmoid (12 steps).
Final Pass is Manual Sigmas only for stability.
Audio volume default set to 0 (neutral) with easy slider control.
Enhanced image conditioning using LTXVImgToVideoInplaceKJ in both passes.
Cleaner workflow layout, better group organization, and more helpful in-workflow tips.
✅ Core Features
Extremely strong, natural and consistent motion from First to Final Pass
Excellent anatomy, lighting, physics and NSFW detail
High-quality 24 FPS video with perfectly synced audio
Easy model switching between 10Eros V1 FP8 and Distilled 22B
Built-in Video Editor (RTX Super Resolution + nmkdSiaxCX upscaler + RIFE 48/60 FPS)
Optimized for 16GB VRAM setups and 32GB RAM
📥 Required Models + Direct Download Links
10Eros V1 FP8 (Main Checkpoint) → ltx2310eros_v1_FP8.safetensors
Distilled Model → ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors
Text Encoder → gemma_3_12B_it_fp8_e4m3fn.safetensors
Text Projection → ltx-2.3_text_projection_bf16.safetensors
Video VAE → LTX23_video_vae_bf16.safetensors
Audio VAE → LTX23_audio_vae_bf16.safetensors
Preview VAE → taeltx2_3.safetensors
Spatial Upscaler → ltx-2.3-spatial-upscaler-x2-1.0.safetensors
Model Upscaler → nmkdSiaxCX_200k.safetensors
Distilled LoRAs (First & Second Pass)
📁 Folder Structure
ComfyUI/
├───📂 models/
│ ├───📂 diffusion_models/
│ │ └─── ltx2310eros_beta.safetensors
│ │ └─── ltx-2.3-22b-distilled_transformer_only_fp8_input_scaled_v3.safetensors
│ │
│ ├───📂 text_encoders/
│ │ └─── gemma_3_12B_it_fp8_e4m3fn.safetensors
│ │
│ ├───📂 clip/
│ │ └─── ltx-2.3_text_projection_bf16.safetensors
│ │
│ ├───📂 VAE/
│ │ └─── LTX23_video_vae_bf16.safetensors
│ │ └─── LTX23_audio_vae_bf16.safetensors
│ │ └─── taeltx2_3.safetensors
│ │
│ ├───📂 latent_upscale_models/
│ │ └─── ltx-2.3-spatial-upscaler-x2-1.1.safetensors
│ │
│ ├───📂 upscale_models/
│ │ └─── nmkdSiaxCX_200k.safetensors
│ │
│ └───📂 Lora/
│ └─── ltx-2.3-22b-distilled-lora-1.1_fro90_ceil72_condsafe.safetensors
│ └─── ltx-2.3-22b-distilled-lora-384-1.1.safetensors
Custom Nodes Overview – MrXin LTX 2.3 I2V EROS V6
ComfyUI-KJNodes
rgthree-comfy
ComfyUI-easy-use
ComfyUI-mxToolkit
ComfyUI-VideoHelperSuite
ComfyUI-LTXVideo
controlaltai-nodes
comfyui_nvidia_rtx_nodes
GACLove/ComfyUI-VFI (voor RIFE)
comfyui_memory_cleanup + comfyui-impact-pack
Pro Tip: Use the new 10Eros V1 FP8 model with its recommended trigger words for the best results.
✅ Quick OOM Fix Guide
If you’re getting Out of Memory errors or weird faces in the final pass, follow these steps in order:
1. Update or Fresh Install ComfyUI + Install Missing Custom Nodes
Download and run the Easy Installer: 👉 https://github.com/Tavris1/ComfyUI-Easy-Install
Or watch this short video: 👉 https://www.youtube.com/watch?v=CgLL5aoEX-s&t=1041s
After updating/installing:
Open ComfyUI → go to Manager → click Install Missing Custom Nodes
Restart ComfyUI completely.
2. Add the Low VRAM Flags
Go to your ComfyUI folder.
Find the file ComfyUI.bat.
Right-click → Edit (open with Notepad).
Find the line that has ComfyUI\main.py .
Add the flags at the end so it looks like this:
bat
ComfyUI\main.py --lowvram --reserve-vram 6 --preview-method none --disable-xformers --disable-smart-memorySave and launch ComfyUI using this .bat file.
3. Increase Windows Swap File (Virtual Memory) This helps when VRAM still spikes:
Right-click This PC → Properties → Advanced system settings.
Click Settings under Performance → Advanced tab → Change (under Virtual memory).
Uncheck “Automatically manage paging file size for all drives”.
Select your main drive (usually C:) → choose Custom size.
Set:
Initial size (MB): 49152
Maximum size (MB): 98304
Click Set → OK → restart your PC.
After doing all three steps, reload the workflow and test with Longer Side 1024 and Video Length 20 seconds first. This combination fixes OOM for almost everyone on 12GB cards.
Version History & Previous Updates
V5 — Major base model upgrade to 10Eros V1 FP8. Significantly better realism, anatomy, lighting and motion coherence. Updated model loading and separate LoRA handling for each pass.
V4 — Introduced separate Manual Sigmas for First and Final Pass, much stronger Final Pass motion, full TwoWaySwitch system for model selection, expanded LoRA stack, and refined Video Editor.
V3 — Solved the biggest V2 complaint by preserving much more motion in the Final Pass. Added Model Upscaler (nmkdSiaxCX), option to disable Final Pass, distilled model support, and custom resolution.
V2 — Added full built-in Video Editor (RTX Super Resolution + RIFE to 48 FPS), improved LoRA suite, better audio sync, and major stability enhancements.
V1 — The original production-ready workflow. Dual-pass system, strong audio integration, excellent low VRAM optimization (12GB cards), and live previews.
Disclaimer:
This workflow is provided for entertainment, artistic, and creative purposes only. It may not be used for any illegal, harmful, non-consensual, or malicious activities. Please use it responsibly and respect all applicable laws and ethical guidelines.
— MrXin (May 2026) 🔥
Description
FAQ
Comments (59)
Hey! First of all, huge thanks for the insanely detailed breakdown — especially the part about which files to download and where to place them. That honestly saved a ton of time 🙏
Also big appreciation for sharing the workflow itself — looks very well thought out. I’ll definitely test it on my side and will drop you a Buzz if it performs as good as it looks.
The video examples are 🔥 btw, really impressive results.
Quick question: since this workflow is optimized for low VRAM, is there any way to take advantage of a high-end GPU like a 5090? Or does this kind of optimization only affect speed and not the final quality?
Hey! Thanks a lot — really appreciate the feedback 🔥
Short answer: yes — with a high-VRAM GPU (like a 5090 or 24GB+), you can get noticeably better quality and faster performance.
You can increase the resolution by disconnecting the Longer Side slider. Then you can manually enter your own resolution. The maximum supported resolution is 4K UHD (3840 on the long side).
You can also download a larger Gemma model:
gemma_3_12B_it_heretic.safetensors
https://huggingface.co/DreamFast/gemma-3-12b-it-heretic/resolve/main/comfyui/gemma_3_12B_it_heretic.safetensors?download=true
The base model (10Eros) is only available in this size. You could try the LTX-2.3 Full version as well, though I’m not sure if it will work:
https://civitai.com/models/2448150/ltx-23?modelVersionId=2752735
This workflow is really optimized for 12GB VRAM + 32GB RAM.
But feel free to experiment with other models as well!
@MrXin Wow, really appreciate the detailed reply — and especially the links, that’s super helpful 🙏
Honestly, that alone deserves some Buzz, just sent you a bit as a thank you!
One more thing I wanted to ask — can I play around with the LTXVChunkFeedForward (4 chunks) setting? From what I understand, it’s also part of the optimization.
Would increasing or changing it make sense with a 5090?
@Honeyphoria The LTXVChunkFeedForward node is purely a memory optimization and has no impact on the final visual quality.
For a 5090: You could set the chunks to 1 or bypass it. With 32GB of VRAM, you don't need to split the processing into smaller pieces. Setting it to 1 will actually make the generation slightly faster because the GPU can process the data in one go without the extra overhead of 'chunking.'
When to increase it? Only if you are pushing the absolute limits (like 4K resolution or very long video sequences) and you happen to run into an 'Out of Memory' error.
Hello Bro,
very clean and good working Workflow. I like it.
I have generate a Video for 20 seconds in 1280. It took me 13 minutes. I have a 4070 Super 12 GB Vram and 32 GB. Its a good time?
I wanna ask, its possible to use it for T2V with this Workflow for example with toggle to change what i need?!
Thanks for ur Work :)
Hello! Thank you for the kind words, I’m glad you like the workflow.
To answer your question: 13 minutes for a 20-second video at 1280 resolution on a 4070 Super is actually a very good time. For comparison, I manage to generate 25-second videos in about 12 minutes, but I am using Sage Attention in ComfyUI. This is a specific optimization that significantly boosts speed for LTX-Video. If you haven't enabled it yet, I highly recommend looking into it to further improve your render times!
Regarding Text-to-Video (T2V): Yes, it is definitely possible with this workflow. This specific 'EROS' version is optimized for Image-to-Video (I2V) using the Start Image input. To use it for T2V, you would essentially need to replace the image input with an 'Empty Latent' node or bypass the image conditioning. You could add a simple 'Switch' or 'Toggle' node (like those from the Impact Pack or KJNodes) to easily swap between a loaded image and a blank latent, allowing you to use the same sampling logic for pure text prompts.
Thanks for the support!
I use sage attention. Ok i think 13 Minutes its a good time. If i use WAN its took me longer :D
I will try to add a toggle and the nodes what u suggested.
Thanks Bro.
So i did any Tests. All with Sage Attention activated in batfile.
1. 1280x 10 seconds(not 20 sry my mistake) . 13 Minutes Finish.
2. 1024x 20 seconds 15 Minutes, final Video compiling OOM -.-
3. 1024x 15 seconds 13 Minutes, Compiling Final Video to finish took next 10 Minutes = 23 Minutes.
4. I changed the VAE Decode for Final Video with VAE Decode(Tile)
1024x 15 Seconds 9,03 Minutes Finish(VAE Decode Tile)
5. 1024 15 seconds 8,59 Minutes Finish (VAE Decode Tile)
So i think its better to change the VAE Decode Node for Final Video with VAE Decode(Tile).
Now it works fine for me. I will try to put Option for T2V with Toggle. Im not an Expert so it took maybe a little bit time to do it :D
I like ur Workflow and i think it will be my Main Workflow if i can expand with T2V, because the Video Result looks better when my main Flow, with Rife VFi. I dont no why :D Mine is a little faster(15 seconds Video so 7 Minutes) but with yours i have better Video Results.
Thanks Bro. Great Work .
@penko1515 Hey, thx for the info. I wil change it and it wil be added in V2. I wil look into it to make a text to video option.
@MrXin ur welcome. Oh that would be nice. Thanks Man :)
to me is just generating the resize image , what i'm doing wrong ?
Hey there! Don't worry, these workflows can be quite complex when you first dive in, so you're definitely not doing anything 'wrong'—it’s usually just a small setting or a missed step. 😊
It sounds like your ComfyUI is successfully handling the first part of the process (the image prep) but hasn't reached the finish line yet. Here are a few friendly tips to help you get that video rolling:
Take a little stroll to the right: My workflow is a bit of a journey! The resized image you're seeing is just the preparation stage. Try scrolling further to the right in your workspace to find the Video Combine node—that’s where the magic final result will appear.
Keep an eye on the 'Command Center': Check the black terminal/console window that runs alongside ComfyUI. If there's a red error message there (like a missing model or a memory warning), it might have paused the workflow right after the resize.
Patience is key: High-quality video generation like LTX takes a bit of time. If the sampler node (the big one in the middle) has a green border, it's still working hard! Give it a few minutes to complete the render.
Check the connections: Just double-check that you've entered a prompt and that all the nodes look 'active' (not greyed out).
Hang in there! Once you get that first video out, it’s totally worth it. Let me know if you’re still stuck and I'll be happy to help you figure it out! 🚀
Thank you for the "hold my hand" instructions. The journey begins.
My weird face, when does it enter the final stage — OOM... hmm
Change in subgraph the VAE Decode for Final Video. Replace it with VAE Decode(Tile). I did it too and helped me alot. It works fine.
Hey!
The OOM usually happens when VRAM gets too tight during the final pass.
I run ComfyUI with these extra flags to help with stability on 12GB cards:
text
--lowvram --cache-none --reserve-vram 6 --preview-method none --disable-xformers --disable-smart-memoryHow to add them:
Find the .exe file to start comfyui in your main ComfyUI folder.
Right-click it → Edit (open with Notepad).
Look for the line that starts with ComfyUI\main.py .
Add the flags at the end of that line, like this:
text
ComfyUI\main.py --lowvram --cache-none --reserve-vram 6 --preview-method none --disable-xformers --disable-smart-memorySave the file and restart ComfyUI using that .bat file.
I also use a custom Windows swap file (virtual memory) set to: Initial size: 49152 MB Maximum size: 98304 MB
This helps a lot when VRAM spikes.
If you still get OOM after adding the flags, try these quick fixes:
Lower the Longer Side slider (try 1024 or even 960 instead of 1280)
Reduce Video Length (try 15–20 seconds instead of 25)
Let me know if it works! 😊
@penko1515 @MrXin Thank you for your advice. At first, using VAE (Tiled) still ran into OOM errors. So I fixed it by overwriting the files with an older version of ComfyUI on top of the one I was using, then updated it again fresh — and it went back to working normally. My issue was probably that some files didn't get updated properly the last time.
@MrXin Could you make a workflow that uses IC-LoRA-Motion-Track-Control to help with NSFW content, something like this (Link) ?
@g1263495582 Hey, i'm glad it works. The Motion Track looks cool. I wil have a look at it when i have the time 😊
I need these to run:
comfyui-custom-scripts (1)
MathExpression|pysssss
danTheMonk/comfyui-int-and-float (1)
IntToFloat
How do I get them, please?
Hey! No worries, those are easy to fix 👍
You’re missing two small custom extensions:
comfyui-custom-scripts (gives you the MathExpression node)
danTheMonk/comfyui-int-and-float (gives you the IntToFloat node)
Best & fastest way to get them:
Update ComfyUI or do a clean install of ComfyUI using the Easy Installer: 👉 https://github.com/Tavris1/ComfyUI-Easy-Install
Follow this quick setup video (it shows the whole process): 👉 https://www.youtube.com/watch?v=CgLL5aoEX-s&t=1041s
After installing, open ComfyUI → click Manager → click Install Missing Custom Nodes. It should automatically find and install the two missing ones.
Restart ComfyUI and everything should work.
Let me know if you still have any issues after that! 😊
@MrXin Thank you.
This has been working great for me, thank you so much. Would it be possible to get an option for first and last frame to vid control?
Hey!
Thanks a lot for the kind words, really glad it's working well for you! 😊
Regarding first and last frame video control — yes, that's a great idea and definitely possible to add.
I'll try to make an updated version with that option. However, my time is pretty limited right now, so it might take a little while. Please be patient :p
Thanks again!
Seconding this wish, please !
Works well but not sure why i need a image as it doesn't use the image except in a brief flash at the start of the video. Am i doing something wrong? I used the prompt that came with the workflow.
Hey!
Thanks for the feedback!
I think the workflow still has missing custom nodes inside the Processing Video subgraph (the big custom node).
Please do this first:
Update or do a clean install of ComfyUI using the Easy Installer (this usually fixes missing nodes automatically):
Download here: https://github.com/Tavris1/ComfyUI-Easy-Install
Watch this short setup video (it shows exactly how to install/update): https://www.youtube.com/watch?v=CgLL5aoEX-s&t=1041s
Open ComfyUI → go to Manager → click Install Missing Custom Nodes.
Restart ComfyUI completely.
It should work after these steps.
I no longer get node errors. I can load an image and have it resized. However, I have not progress past that. I may not understand the QUEUE PROMPT which I believe to be simply "r".
Hey!
Thanks for the feedback!
I think the workflow still has missing custom nodes inside the Processing Video subgraph (the big custom node).
Please do this first:
Update or do a clean install of ComfyUI using the Easy Installer (this usually fixes missing nodes automatically):
Download here: https://github.com/Tavris1/ComfyUI-Easy-Install
Watch this short setup video (it shows exactly how to install/update): https://www.youtube.com/watch?v=CgLL5aoEX-s&t=1041s
Open ComfyUI → go to Manager → click Install Missing Custom Nodes.
Restart ComfyUI completely.
It should work after these steps.
Everything is clear! The women are moaning nonstop.
Hey!
Thanks for trying the workflow 👍
When you say “after previewing the image, nothing happens”, it usually means some custom nodes inside the big Processing Video subgraph are still missing.
Please do this first (best solution):
Update or Fresh Install ComfyUI
Download and run the Easy Installer: 👉 https://github.com/Tavris1/ComfyUI-Easy-Install
Or watch this short video for the full process: 👉 https://www.youtube.com/watch?v=CgLL5aoEX-s&t=1041s
After updating/installing:
Open ComfyUI → go to Manager → click Install Missing Custom Nodes
Restart ComfyUI completely.
Once you’ve done that, reload the workflow and try queuing again. The previews should start working properly.
Let me know if it still doesn’t move after restarting — we can check the next possible cause.
Thanks! 😊
@MrXin Yes, I figured it out. For some reason, you put the model in checkpoints instead of diffusion models. Everything worked, thanks.
@Renessance Glad to hear you got it working! 😊
Yeah, the Text Encoder (gemma) extracts the CLIP from the base model, so it needs the checkpoint to be placed in the checkpoints folder (not diffusion_models).
Enjoy the generations and feel free to tag me if you make something cool! 🔥
@MrXin If you add an RTX super resolution node there, it will be absolutely super.
@Renessance Yeah you're right, but not everyone is able to run that node.
@MrXin If you'd like, please send me a screenshot showing where to insert the node and how to connect it. I'd be grateful, thanks.
Do you keep dr34m and furry lora at .50 for the most part?
Hey, yeah indeed. For me its the sweetspot for most of the LTX 2.3 lora's.
I am trying this for the first time right now and I noticed the following message in the console:
"ERROR lora diffusion_model.audio_adaln_single.linear.weight shape '[18432, 2048]' is invalid for input of size 25165824".
I am fairly new to all this, so I am not sure what to make of that message. Any insight would be greatly appreciated!
Hey!
No worries i'm going to help you👍
What the error means:
The message ERROR lora diffusion_model.audio_adaln_single.linear.weight shape '[18432, 2048]' is invalid for input of size 25165824
means one of the LoRAs you have enabled doesn’t match the exact checkpoint (ltx2310eros_beta.safetensors). It’s trying to load audio-related weights that have the wrong shape.
Quick fix :
In the workflow, find the node called “Load Lora’s” (the big rgthree Power Lora Loader with the 4 lora's in it).
Turn OFF all four LoRAs or any other lora's you may have added:
Set every “on” switch to false (or set strength to 0)
Leave the Distilled LoRA node (the one below it) ON — that one is safe.
Hit Queue Prompt again.
This almost always fixes the error immediately.
After it works, you can turn the extra LoRAs back on one by one to see which one was causing the problem (most of the time it’s either the NSFW furry one or the cr3ampi3 animation one).
Let me know if it works or if you still see the error after disabling them! 😊
Is this something that happens with 2.0 loras or 2.3 loras?
@tenstrip it could be. the Lora's from LTX 2 can work with LTX 2.3 but sometimes not. Try to test it with only the DR34ML4Y - All-In-One NSFW lora. This one works fine.
Hi! I’ve been testing your workflow and ran into an issue I can’t quite figure out.
After around the 10-second mark, noticeable artifacts start appearing on the character’s face and around it. What’s interesting is that the video preview after the second pass looks clean — the problem only shows up after the final pass.
I’m currently using the Spatial Upscaler v1.1, so I’m wondering if that might be causing the issue. I also tried adjusting CFG, sigma, and strength values in the final pass, but unfortunately it didn’t help.
Have you encountered something like this before, or do you have any ideas what might be causing it?
Thanks in advance!
Hey!
Sorry to hear you're getting those artifacts after the 10-second mark 😕
The fact that the second pass looks clean but the problem only appears after the final pass is a good clue. It’s likely related to the spatial upscaler or the re-injection step in the last pass.
I’ve only tested with the official V1.0 of the Spatial Upscaler so far. I’ll download V1.1 today and test it myself to see if it causes the same issue.
I’m also currently working on V2 of the workflow, so I’ll make sure this kind of artifact is fixed or at least much better handled there.
Thanks a lot for the detailed report and info — really helpful!
I’ll let you know once I’ve tested V1.1 or when V2 is ready 👍
@MrXin I'm also gonna test with v1.0 spatial today, looks like it can be only the reason. I will give you a feedback bro. I should’ve tested it first before writing — sorry about that 🙏
It would also be really cool if you could add Tensor or RTX upscaler and RIFE interpolation to the new workflow, along with a switcher to quickly turn them on and off. That would be super useful.
Overall though, you made a really great workflow — thank you for that!! 🙌
Hey!
I tested the Spatial Upscaler v1.1 and it works fine on my end — no artifacts.
The issue is probably caused by the VAE Decode (Tiled) node. There are two different nodes:
Normal VAE Decode
VAE Decode (Tiled)
I use the normal VAE Decode (not the tiled version) in the final pass. When I switch to VAE Decode (Tiled) I also get similar face artifacts after ~10 seconds.
Try switching back to the regular VAE Decode node (disable or remove the tiled one) and queue again. That usually fixes it.
Let me know if that solves the artifacts for you! 😊
@MrXin Hi bro! Looks like i have already normal VAE decode node in your workflow by default. Here's a screenshot. You sad you fixed the issue with replacing vae decode node, can you help me please 🙏🏻 Also i bypass LTXV Chunk FeedForward because i don't need it.
Great workflow. I am testing it out now. It generates the first pass, but ends there and does not go father, like the workflow is done. How can I diagnose it? Thanks!
Hey!
Thanks, glad you like the workflow! 👍
If it only generates the First Pass and then stops (as if the workflow is finished), it usually means a custom node is missing or not loaded properly in the later passes (most often the upscaler or one of the nodes inside the big Processing Video subgraph).
Best way to fix this:
Update ComfyUI (or do a fresh install) 👉 https://github.com/Tavris1/ComfyUI-Easy-Install
After updating:
Open ComfyUI → go to Manager → click Install Missing Custom Nodes
Restart ComfyUI completely
Reload the workflow and try again.
This solves the problem for most people who only get the first pass.
Let me know if it still stops after the first pass after you’ve done this — then we can look deeper (for example if the upscaler model is correctly loaded).
Good luck!
@MrXin Thanks for the fast reply! I forgot there are logs, so I checked and noticed one model was not loading, likely causing the issue (different folder name). Doing another run now! Surprised it didn't throw an error.
@Shnoof No problem at all! Glad you found the issue 😊
ComfyUI is sometimes pretty buggy.
Anyway, hope the new run works perfectly now!
Let me know how it turns out 🔥
running this on a 3090 and it works fine. takes some time but it works.
Hey!
Glad to hear it's working well on your 3090! 🔥
Out of curiosity: How long does it take you to generate one full video, and what settings are you using? (Longer Side + Video Length)
For reference, on my RTX 5080 12GB card it takes me about 12 minutes with Longer Side 1280 and 25 seconds length.
Would love to know the speed on a 3090 😊
@MrXin It takes 18 minutes on "default" settings. 1280,20 seconds. and around 7-8 minutes for 1280,10 seconds. I prefer generating 6-7 second clips for realistic content because the drift in the likeness becomes severe.
I tried the fresh install as you suggested and disabled the loras except for dreamly and it still gets blurry / grainyer with each pass
Hey!
Sorry to hear your video is getting blurry/grainy as it progresses 😕
This usually happens when one (or more) of the required models is missing, incorrectly placed, or the wrong version.
Please double-check the following:
You have all the correct models downloaded:
ltx2310eros_beta.safetensors → checkpoints folder
gemma_3_12B_it_heretic_fp8_e4m3fn.safetensors → clip folder
taeltx2_3.safetensors → VAE folder
ltx-2.3-spatial-upscaler-x2-1.0.safetensors → latent_upscale_models folder
ltx-2.3-22b-distilled-lora-dynamic_fro09_avg_rank_105_bf16.safetensors → Lora folder
Make sure they are in the exact folders as shown in the workflow description.
After placing the files, restart ComfyUI completely and try again.
A lot of people get grainy/blurry results when the spatial upscaler or the preview VAE is missing or in the wrong folder.
Let me know if everything was in the right place or if it’s still blurry after checking — then we can look at other possible causes.
@MrXin You fucking rock