🚀 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
🚀 V2 – What’s New & Improved (March 2026)
Built-in VIDEO EDITOR (bypassable) → fully automatic 48 FPS final video
RTX Video Super Resolution (1.5x + ULTRA quality)
RIFE Interpolation (2x FPS)
Two-stage workflow (generate base → then upscale) to keep VRAM usage low
Newer & better models
Spatial Upscaler upgraded to v1.1 (ltx-2.3-spatial-upscaler-x2-1.1)
Lighter & faster text encoder (gemma_3_12B_it_fp8_e4m3fn)
Expanded LoRA suite (more explicit power)
Added: ltxdeepthroat_v01, Penis Lora, liquid_wet_dr1pp, LTX-2.3 - Orgasm, LTX2.3TITFUCKE2000 + stronger DR34ML4Y & cr3ampi3
Improved final decode with LTXV Spatio-Temporal Tiled VAE Decode (sharper details, fewer artifacts)
New trigger-word notes added directly in the workflow (Deepthroat, Creampie animation, Stomach bulge, Wet Dr1pp, etc.)
Cleaner output folders (/I2V/ for base + /Editor/ for 48 FPS final)
Extra stability & VRAM optimizations (VRAM_Debug, Fast Groups Bypasser, better batch management)
Result: Same rock-solid 12GB VRAM performance, but noticeably smoother motion, much better explicit details, higher final FPS (48) and overall higher visual quality.
V1 → V2 upgrade is 100% worth it if you want that extra polish and modern 48 FPS look without extra hassle.
FAQ
Comments (87)
Hey bro, huge thanks for the workflow! 🙏 I sent you 2000 Buzz — honestly you deserve like 200,000 Buzz 😄
I’ve got a few questions if you don’t mind:
I’m running on an RTX 5090, planning to generate with FP8 model + FP8 text encoder, and resolution 1280. From what I understand, I can fully disable LTXV Chunk FeedForward — is that correct?
My main question is about the new LTXV Spatio Temporal Tiled VAE Decode node. Since I don’t really need VRAM optimization, what settings would be optimal for my setup?
Also, should I keep the switches inside the VRAM Debug node enabled, or can I safely disable them?
And one more thing — about Sigma values:
For a 20-second video, the recommended Sigma is 0.9175, while for 15 seconds it’s 0.9295. Shouldn’t Sigma be higher for longer videos instead of lower?
Thanks again for the awesome workflow 🙌
Also bro it would be cool to add Tensor upscaler because as i figured out seems like RTX Upscaler doesn't work on cloud servers. I'm using comfy on cloud
Hey bro! 🔥
First of all — massive thanks for the 2000 Buzz and the super kind words, really appreciate it man 🙏 Made my day!
Now to your questions:
✅ LTXV Chunk FeedForward Yes, you can completely disable it (or even delete the node). On an RTX 5090 you have tons of VRAM, so the chunking is no longer needed. Feel free to turn it off.
✅ LTXV Spatio Temporal Tiled VAE Decode Since you don’t need VRAM optimization, you can replace the entire Tiled VAE Decode node with the normal VAEDecode node. It’s faster and usually gives cleaner results when you have enough VRAM. Just connect the latent directly to a regular VAEDecode.
✅ VRAM Debug node You can safely disable all the switches inside the VRAM Debug node (set empty_cache, gc_collect and unload_all_models to False). It’s only there for 12GB cards.
✅ Sigma values Good question! I tested a lot of different sigma combinations. Even though it feels counter-intuitive, lower sigma for longer videos (0.9875, 0.87, 0.735, 0.445, 0.0) gave me the most stable and best-looking results with this model. Other sigma on longer videos often made the motion more chaotic or added weird artifacts. So I kept the values that worked best in practice.
If you want to experiment, feel free — but these settings are currently my personal “sweet spot” for quality and stability.
Let me know how it goes on the 5090, I’m curious how fast it runs for you 😄
Any other questions, just hit me up!
— MrXin
Hey bro! 👊
Thanks for the feedback.
Yeah, I understand — the RTX Video Super Resolution node only works on local machines with a real NVIDIA RTX GPU. It doesn’t work on most cloud servers (RunPod, Vast.ai, Massed Compute, etc.).
I’ll definitely look into adding a Tensor Upscaler (or a normal model-based upscaler) as an alternative for cloud users. It might take me a bit of time to test and integrate it properly, but I’ll add it in a future update.
In the meantime, you can try bypassing the RTX Upscaler + RIFE part and just use the base 24 FPS output, or replace it with a regular Upscale Model node if you want.
Appreciate you letting me know — this kind of feedback helps a lot!
Let me know if you need any temporary workaround.
— MrXin
@MrXin Thank you! Based on my experience Tensort upscaler provides best spee/quality balance
I'm having trouble getting the RTX Video Super resolution node... Assistance please?
If you have comfy manager just click on it and then click install missing nodes. Then chose all missing nodes click install and after installation click restart
@Honeyphoria Tried that and it still gives me the error. Would updating the whole ComfyUI help this?
@lanceshocker Weird, worked for me😕
Hi! 👋
Thanks for reaching out. I'd be happy to help you with the RTX Video Super Resolution node.
To better assist you, could you please tell me:
What exact error message or error code are you seeing?
Does the error happen right when you load the workflow, or only when you hit Queue Prompt?
Common issues with the RTX Video Super Resolution node are usually related to:
Missing or outdated NVIDIA RTX nodes / custom nodes
Not having an RTX GPU (it only works on NVIDIA RTX cards)
Driver version problems
Once I have the exact error message, I can give you a precise fix.
Looking forward to your reply! I'll help you get it working 😊
— MrXin
@MrXin It's installed, just the workflow is not detecting it. It happens when I load the workflow. I have the latest update but it still shows up red on the workflow. RTX 5080 if this is helpful.
@lanceshocker Hey! 👋
Thanks for the extra info — RTX 5080 is useful to know.
The fact that the RTXVideoSuperResolution node turns red when you load the workflow usually means one of these two things:
The custom node comfyui_nvidia_rtx_nodes is not installed correctly, or
ComfyUI didn’t properly register the node after the update.
Quick fixes to try right now:
Restart ComfyUI completely
Close ComfyUI fully
Restart it (preferably using your ComfyUI.bat)
Re-install the RTX nodes
Go to ComfyUI Manager
Search for comfyui_nvidia_rtx_nodes
Click Install (even if it says it’s already installed)
Restart ComfyUI again after installation
Update everything
In Manager, click Update All
Then click Install Missing Custom Nodes
Restart once more
After restarting, load the workflow again and check if the node turns green.
If it’s still red after doing the above, please try this:
Go to the ComfyUI console window (the black terminal)
Copy and paste here any red error messages you see when the workflow loads.
Also, just to confirm: Are you running ComfyUI locally on your RTX 5080, or on a cloud server?
Let me know what happens after you try steps 1–3. We’ll get it fixed 👍
— MrXin
@lanceshocker Run this command inside your ComfyUI folder (the folder where you can see the python_embeded folder): ".\python_embeded\python.exe -m pip install -U --no-build-isolation nvidia-vfx --index-url https://pypi.nvidia.com" If you run into any problems, take a screenshot of the error + the command line output and send it to Claude/Gemini for help.
@lanceshocker I also encountered this problem on ComfyUI Desktop.
Here's the solution:
1. In the ComfyUI terminal, cd custom_nodes\Nvidia_RTX_Nodes_ComfyUI
2. pip install -r requirements.txt
Hellow! Error - TextGenerateLTX2Prompt - NotImplementedError: Cannot copy out of meta tensor; no data! what is it? help please
Hey! 👋
In my V2 workflow (MrXin LTX 2.3 I2V EROS V2), there is no TextGenerateLTX2Prompt node. I only use the normal CLIP Text Encode nodes for the positive and negative prompts.
This error usually appears when someone loads an older workflow or accidentally adds that node from another LTX setup.
What you can try:
Make sure you're loading my V2 JSON specifically.
Search the workflow (Ctrl + F) for “TextGenerateLTX2Prompt” and delete it if you find it.
Best solution: Do a fresh ComfyUI install:
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 reinstalling, update all custom nodes in the Manager and restart ComfyUI.
Let me know if the error still appears after that — I’ll help you fix it step by step!
— MrXin
Hey, if it's not a secret, how do you generate prompts? Qwen? Grok? On your own?
Hey! 😊
Good question!
I generate most of my prompts using Grok (from xAI).
I give Grok a basic scene description or reference image idea, and then I work together with him to refine and expand it into a detailed, cinematic prompt that works well with LTX 2.3. I usually do several back-and-forth iterations until the prompt has the right timing, camera movement, emotion, and explicit details I’m looking for.
So yeah — Grok is my main helper for prompt creation nowadays. He’s really good at understanding what LTX needs.
Would you like me to show you how I typically prompt Grok to create these videos? I can share my prompt template if you want!
— MrXin
@MrXin I would be very grateful :)
@nailywhite172 this is what i prompt to grok, i also add the refrance image to grok:
this is a exemple prompt from diffrent video:
use this prompt as guide for the my prompt based on my image. also take other sex lines for the woman to say.
"A beautiful elegant young female Japanese idol flight attendant in a fitted grey jacket, blue stripe neckerchief, french bun hair, pretty Japanese face with natural makeups, and a nameplate. (0:00) She is already rapidly thrusting her breasts up and down quickly, giving a passionate titfuck to the passenger's penis in the dimly lit galley corner during a nighttime luxury long-haul flight. Wet skin collision sound and airplane jet noise can be heard. She presses her soft large breasts firmly together and is sliding them up and down rapidly in repeated deep strokes. (0:05) Gentle, sweet, professional yet teasing voice in fluent English. She maintains the consistent fast rhythm titfuck. She looks up with a soft smile and asks "You’re throbbing so hard between my tits, sir~" (0:08) She maintains the consistent fast rhythm titfuck, breasts gliding smoothly and continuously. She keeps eye contact, remaining in the same facial expression. (0:10) The titfuck accelerates further, her breasts bouncing more vigorously and rapidly. A man groans "nngh…" in low masculine voice, and the penis throbs intensely between her cleavage, ejaculating large amount of warm sticky cum that spurts out in several thick ropes from her cleavage, cum sticking on her breasts, dripping on her upper chest. (0:15) She maintains the motion briefly then slows gently, opens her eyes wider with a loving smile, and whispers sweetly "Look at all this thick cum you shot on me… Mmm, naughty passenger♪ Mmm..." She stares at us seductively. A little pause at the end."
One more question. Where do you get the starting image? Is it generated? Or is it just something random from Google? :)
@nailywhite172 Hey, i make them my self with illustrious model
Hey Bro,
ist looks like better than V1. Thanks for ur hard Work and share to us ur Workflow.
I tested it, the first preview pass its ok, i see the action and seems good, but it continue to do the second pass, the action change it, it will repleaced and that most are bad anatomy. wrong, position, not doing what the prompt said. I hoped the second pass is just for finetuning but it change all the action with distorded anatomy. i tried different seed but its same, the second pass ruined the action before, that looked better. Maybe my System are fault i dont no or u have a suggestion what can i change? i have 4070 super 12gb vram 32gb ram. Video in 960x 24 frames 15 seconds. 11 Minutes finish.
Hey, i just did a test with your video settings and it works fine on my end. Whats the resolution height and width of your input image? What upscaler model are you using?
@MrXin Thanks for ur answer. Images are in 1280x1024. I think, i solved it by reduce the LTXVImgToVideoInplace (Second Pass) to 0.6 than with 1.0. The change from 1st pass to 2nd pass are quite similar, anatomy is better. it seems that works now better. I need to do any tests :)
Thanks :)
@penko1515 i'm glad to hear that 😄
It would be good if you could add from the initial frame to the final frame.
Hey, i'm working right now on a text to video workflow. after thats i wil look into it. :)
you can do it by your self not hard. just 5min
This workflow is pretty good, thanks. I'd really like a way to separate the audio out to quash the stupid moaning though.
Hey, you could try another vea audio. and load it with a vae loader.
@MrXin Thanks !
Really nice workflow! I've been messing around with it the past couple days and really like the audio incorporation. Is there any way to like train a lora to get more of a particular voice and add that to the workflow? Not too familiar with the audio side of things
Hey! 👋
Thanks! Glad you like the audio.
Training a LoRA specifically for a voice in LTX 2.3 is possible but quite difficult and not well supported yet.
The easiest way right now is:
Generate the video with placeholder audio
Create the voice separately using XTTS or RVC
Then combine the video + custom audio afterwards.
— MrXin
@MrXin Coming back around to this thread. I actually saw there's a newer node called "LTXV Reference Audio (ID-LoRa)" that might work for this exact idea. I've been trying to implement it for the workflow to change the sound of voices and haven't had any luck. Do you think you could try and see if you could successfully implement it into your workflow? I think a lot of other people would be interested in that too. Thank you!
Very nice workflow, easy to understand and start using LTX. Coming from WAN there are a few differences here and there but this flow makes it easy to get started.
I'm still trying to figure out prompting but wonder if you can share any advice - the first pass preview generally looks good with strong movement and detail (though some occasional rough edges) but the final pass often deletes ~75% of the movement for me. Is there something I should tweak for this?
Thanks again for the great resource, looking forward to your future projects
Hey! 👋
Thanks a lot, glad you're enjoying the workflow!
The issue you're describing (strong movement in first pass, but much less in final pass) is pretty common with the spatial upscaler.
Try these two tweaks:
In the Final Pass, find the node LTXVImgToVideoInplace and lower the strength to 0.6 (or even 0.55). This usually helps keep more movement.
You can also experiment with the ManualSigmas in the Final Pass. Try these values:
text
0.87, 0.735, 0.445, 0.0Let me know if that improves the motion in the final video!
Happy generating!
— MrXin
@MrXin Thanks for your very helpful reply.
The LTXVImgToVideoInplace node only seems to accept 1 decimal place (0.5 and 0.6 were ok, but wouldn't take 0.55). Changing the default 1.0 to 0.6 didn't seem to have much of a difference, but the change to ManualSigmas seems to have helped a lot. It looks like there are many more things to tweak with LTX compared to WAN, I'll have a busy week ahead figuring it out.
Thanks again!
It's strange, but Kijai's workflow is 1.5 times faster. I can't figure out why it's slowing down on yours.
hi, where can i find his workflow?
@Axsee In the KJ workflow, the preview sampling runs at 8 FPS, whereas this one is at 24 FPS. Also, after the first stage is finished, there's an additional decode process just for the preview output.
IDK bro, check please this video. It has some weird white aura around character and same face artifacts even with V2 workflow. What can i do possible wrong? Everything turned by default on workflow except butch node. Thank you for the answer. I hope it will be no meaningless AI auto answer😅
Yeah, i know this is a problem. I'm trying to find a fix. But its not easy. Its mostly the sigmas and the VAE decode part.
@MrXin Got it, I thought maybe I was doing something wrong on my side. Good luck to you! You’re doing something really big. I believe that over time we’ll move from Wan to LTX.
I want you to understand — right now you’re building the best LTX workflow in the world. No exaggeration. It’s not perfect yet, but it’s definitely the best. Maybe at this point the issue isn’t even the workflow itself anymore, but I really hope you’ll find solutions to the current issues.
Why do we use the checkpoint in everywhere like vae textencoder ? is it baked in checkpoint?
yes, indeed. I havent test it with other vae textencoder.
Hey guys if you get this : SamplerCustomAdvanced
The size of tensor a (19760) must match the size of tensor b (2609536) at non-singleton dimension 2.
-> Then the FIX is to DELETE the ComfyUI_smZNodes folder from the custom nodes folder. The mere presence of that folder breaks the sampler and furthermore it's not even required for this workflow.
Hey! 👋
First of all, sorry you lost 3 hours on this — that sucks.
Answer to your questions:
ComfyUI_smZNodes → No, it is NOT required for this workflow (V1 or V2). This workflow only uses standard LTX nodes + a few common ones (rgthree, KJNodes, easy-use, etc.). smZNodes was conflicting with the SamplerCustomAdvanced node and causing the tensor size error. Deleting that folder was the correct fix.
Only getting 5 seconds instead of 25 seconds This is not caused by deleting smZNodes. The video length is controlled by the "Video Length (seconds)" slider (currently set to 25).
Quick fix for the short video:
Make sure the Video Length slider is really set to 25
Check that the connection from the slider → the subgraph (the big "Processing Video" group) is still there
Again, sorry for the headache — happy to get this sorted for you.
— MrXin
Works very fine finally ! Very good workflow
Please tell me, you have processing, first pass and final video in your workflow, but I like the result in first pass the most, it is the movement that is more energetic, and in the final video it is already slower, why is that? how to get the result from first pass?
Hey, i'm aware of the problem. I'm looking into it right now. Its has something to do with sigmas. But its not easy to find the right balans.
@MrXin in fact, I just turned off the upscaler, lowered the quality a little, and it looks like the result is much better, but not perfect. In some poses in the image, the man refuses to move at all.
Great workflow, i just tip you 400 buzz (sorry i don't have much). but honestly this is the best LTX workflow i have ever used.
Just one question i'm using wf v2 and sometime the action of the video is less than the first pass preview, i read some comment about this issue, some say use the ManualSigma in the LTXvimagetovideoinplace but i coudn't find that node in both wf v1 and v2
Could you please show me how to do that?
Thanks for your great work!
I have the same thing, the preliminary video seems to be better than the final result, I want to focus on the preliminary one, yes, you can download it, but the quality is poor, I also tried different things, I don't know how to do it, I hope the author will help))
@baamad9400 Hey, i'm aware of the problem. It's related to the upscaler, i need to find the right sigmas for it, so the motion is kept and the image is still good. Its really tricky.
@deatknight0415164866
thx for the buzz and the support, really i appreciated it.
Thanks for your reply
btw i found the ManualSigma node in the final pass
I have try some value and look like it work
But there is another problem is that even my input image is realistic style it render to 3d character video
I don't know is this the LTX 2.3 issue or i have to instruct in my promt
i was using the prompt like this
this is my video https://civitai.com/posts/27626062
[Scene] Edge of Olympus with storm clouds swirling and marble cracking, glowing divine energy auras, lightning arcs and golden sparks, epic and chaotic fantasy atmosphere.
[Characters] Hera (mature regal, long voluminous dark raven-black wavy hair, piercing icy hazel eyes, voluptuous body, wearing sexy sheer translucent silk). Aphrodite (youthful golden-blonde wavy hair, violet-blue eyes, extremely voluptuous, wearing sexy sheer translucent silk). Athena (athletic dark chestnut brown hair, steel-gray eyes, toned curvy body, wearing sexy sheer translucent silk).
Shot 1 (Epic Wide Dramatic, 4s): Low-angle dramatic pull-back with slight Dutch tilt. Three goddesses in foreground with outstretched hands, silk billowing wildly, energy crackling between them.
Shot 2 (Medium Power Pose, 3s): Camera circles the trio as lightning and sparks intensify, marble cracking beneath their feet.
Shot 3 (Final Wide Freeze, 3s): Deep resonant male narrator voice booms: "One throne. Three goddesses. A war written in silk and starlight. Who will rule Olympus?" Final swell of epic orchestral trailer music with massive choral hit, pounding drums, and fading thunder.
This is awesome. Thanks for sharing this!
Two questions, if you do not mind me asking :-)
1. Concerns the second part of the workflow. It gives me this error: "Node 'ID #130' has no class_type. The workflow may be corrupted or a custom node is missing.". That's the RTX node, which I do have installed. When I bypass this part, the rest works perfectly fine. Any clue?
2. Your workflow mentions a "p*nis ltx23" and a "liquid fluid ltx23" lora. Both no longer seem available. ANyone any clue where to get these perhaps?
Thanks! I will tip you the few buzz I have :-)
hey, this is a answer from another user. Maybe this will help.
Run this command inside your ComfyUI folder (the folder where you can see the python_embeded folder): ".\python_embeded\python.exe -m pip install -U --no-build-isolation nvidia-vfx --index-url https://pypi.nvidia.com" If you run into any problems, take a screenshot of the error + the command line output and send it to Claude/Gemini for help.
@MrXin Hi appreciated the help! Installing it is not the problem, I actually already had installed it manually. It just complains about the node.. Not sure, I'll look into it a bit more tomorrow :) Any change you could give some advice on those two loras perhaps btw? :) THank you!
@whiteduck590 this is for the liquid: https://civitai.com/models/2348942/ltx-2-liquid-physics?modelVersionId=2644206
and the penis one is gone from civit :(
@MrXin Thanks! Should anyone have the penis ltx23 lora somewhere on another host or if someone could share it, highly appreciated :)
@whiteduck590 You can try PENIS LoRA by CoachBate - LTX-2 - LTX-2 v3 | LTXV2 LoRA | Civitai
It keeps penises pretty consistent, but the only caveat is it will mess with the audio generation unless you route the model pipe to an "LTX LoRA Loader Advanced", load it in that node, and turn the video_to_audio, audio and audio_to_video settings all the way down to 0. I usually put it between the power lora loader and the load distilled lora in the workflow.
@ArcleinSK thanks will try!
Hello, i have a question. What does cfg do? Since i want my character to do exactly what i wrote in my prompt. Like if i want my character to sleep and do nothing, the video makes the character start moaning and move around. I even added "moving character" or even motion in negative prompt but doesn't stop the character from moving and moaning. Also i disabled all the loras.
I would appreciate it so much if i can get some guidance to make my character stop moving and just have a nice sweet dream (aka sleeping and not moving around with moans.) Also to stop the moans too if possible, even though i typed moan in negative but it keeps doing it.
Hello, I really like this, work so well. The preview really helpful too. Can you guide me how to make this T2V instead of I2V?
Hey, i'm currently working on it. It wil be posted in the near future.
LTXVImgToVideoInplace has a bypass toggle. When on, it will ignore the conditioning coming from an image input. So it's basically 1.00 denoise. You can also lower the strength to get a partial influence. The latest WF that I posted has I2V and T2V versions if you want to look at the difference in the latent chain. It's really very little that needs to be changed. You can still use the image input resizer to set your latent size. Sigmas will need to be appropriate for T2V too, so pick a different schedule.
this workflow gives me nothing but blurry messes.
Hi! Still been using the workflow and getting videos out just fine, but noticed these errors in the console when first running the prompt:
got prompt
Failed to validate prompt for output 161:
* VRAMCleanup 163:
- Required input is missing: input
- Value not in list: cleanup_mode: 'True' not in ['Cache Only', 'Text Encoder', 'Full Cleanup', 'T2V + QwenVL Fix']
Output will be ignored
Failed to validate prompt for output 163:
* (prompt):
- Required input is missing: input
- Value not in list: cleanup_mode: 'True' not in ['Cache Only', 'Text Encoder', 'Full Cleanup', 'T2V + QwenVL Fix']
Output will be ignored
Not sure if this means the VRAM cleanup hasn't actually been running for me or not. It takes around an hour to get a 25 second video w/ audio at 1024 longest side resolution (RTX 3060 12gb w/ 32gb sys ram). Genuinely 5x faster than it would've taken for WAN so not a big deal at all, just curious if I could be getting more speed.
Are some of the loras missing from the description?
Edit:The models all seem to have different names depending on where they were mentioned, but the ones that don't have any links seem to be -
cr3ampi3/Creampie animation, liquid_wet_dr1pp/Wet Dr1pp, Penis Lora, Stomach bulge
Something I discovered while testing - the preview on the final pass inside the "Processing Video" node group seems to spike VRAM usage and can lead to OOMs. I patched this locally by adding a try/except around the node's self.decode_latent_to_preview call so all we lose is a preview frame rather than the entire generation. Those previews seem pretty aggressive
Heretic works so much better than abliterated. I heartily endorse your choice. It's easy to not notice when gemma is strangling your prompts- it can kind of look like it's adhering, but it's just your LoRA bleeding through. Just an FYI for anyone having trouble with LTX promptage.
reading your comment i noticed that heretic was only listed for v1.0 of this workflow but v2.0 uses a non-heretic version, i wonder why this is
downloaded both and i'll compare them, thanks for posting about heretic i'd never even heard of it
I wouldn't use either, just the default. I haven't found any situations yet for seed v.s. seed where the abliterated gemma did anything better. Most of time it's the same with worse dialogue.
@tenstrip So you put it in the required models list so people wouldn't use it? Very clever.
@seaborgiumthemad I made the comment as a general statement on heretic, it's powerful when used with the vanilla model...the OP's point being that a merged model has a better chance of working with a standard encoder, as the nsfw stuff is baked in. That's a fair point, though I don't entirely agree. When I run gemma through ollama and look at exactly what it is letting through, I see a lot of stuff just being ignored completely. Abliterated loosens the leash a bit, heretic lets it all fly. OP is right in that if you can get away with the regular encoder you'll have a better chance of a cleaner image and the model will pick up the slack. I like heretic, you don't have to write a novel-length prompt to tease what you want out of the encoder. To each his own.
@Ponder_Stibbons I didn't make this, it's definitely not required you could just use the fp8 scaled gemma or q4 one.
Holy shit it works.
Question for you. I notice that the 1st pass preview has a lot of expression and motion and emotion that looks really promising, then the final video is super dulled down and way less expressive and has less motion and expression... Is that a me problem with my settings, or is there some way to fix this? Thanks!
Just right click the width and height node and go to the properties then adjust the maximum.
How can you add audio to a wan 2.2 gen base video?