All credit goes to:
Phr00t / LTX2‑Rapid‑Merges and Lightricks / LTX‑2
I simply combined their workflows into a single pipeline that supports Image‑to‑Video, RIFE interpolation, and Latent Upscaling.
The workflow includes NSFW models; however, SFW alternatives are available on Phr00t’s Hugging Face page.
Model Placement
Diffusion model:
\diffusion_modelsVideo VAE:
\vaeAudio VAE:
\checkpointsText encoder:
\text_encodersUpscaler:
\latent_upscale_models
Edit: Feel free to delete any custom nodes in the Optional Group if you don't need them.
V3.0 Update Notes
Bug Fixes: Resolved issues by adding the missing
LTXVCropGuidesnode. If your customized workflow is glitching, adding this node will noticeably improve stability.Automation: Added automatic aspect ratio detection and automatic dimension rounding to multiples of 64.
LoRA Support: A LoRA chain has been added as an example inside the Samples section.
Crucial: Ensure both samplers use the exact same internal configuration; only the external values should differ. This is necessary to avoid visual artifacts.
Final Note: This may be my last update for this project. Thanks for the support!
V2.5 Update Notes Disclaimer: This is essentially the original Lightricks workflow. I’ve worked hard to find more reliable settings, but this is the best result so far, I've mainly removed dependencies and reorganized the nodes.
Most artifacts are caused by the Strength settings, the prompt, or the source image. As shown in the video example, glitches don't occur every time, but some images are more prone to them than others. Here are some tips to avoid these issues:
Adjust Strength: Avoid using full strength. Experiment with different values for Stage 1 and Stage 2; every LoRA requires specific tweaking.
Tweak Compression: For high-motion videos, try increasing the compression. If the video looks stable, you can lower it to improve sharpness.
Frame Rate: Generally, a higher frame rate leads to better stability.
Resolution & Upscaling: Note that the Resolution and Frame Rate settings represent the processing values. The final video output will be twice the size of these inputs.
Final Troubleshooting: If the above tips don't work, try changing the prompt, the seed, or as a last resort, the source image itself.
The workflow includes the wrong link to optionally download Gemma, this is the correct link:
V2.0 Update Notes
LoRA Support & Resolution: This version adds LoRA support, optimized for 720p. While I aimed for 1080p, 720p currently provides the most stable results and better visual coherence when using LoRAs.
What to expect:
Better Consistency: LoRA details are more reliable at this resolution.
Fluidity: The 8-sigma process is fast and works great for standard motion.
Audio Sync: 720p ensures the best synchronization between audio and video.
Note: This update is the result of many trials to balance quality and stability. I’ll keep working on a high-resolution version for future releases.
Edit: update sigmas from 8 to 12 as 1.0, 0.98, 0.96, 0.94, 0.91, 0.87, 0.82, 0.75, 0.55, 0.35, 0.15, 0.0
Performance
Rendering a 5‑second video 1024x1400@30fps takes approximately 2 minutes on a system with 16 GB of VRAM.
Licensing & Usage
All usage rights, licenses, and permissions are governed by the original authors.
Please refer to the official source pages and repositories for the applicable terms and conditions.
For detailed instructions and model usage, please consult the original workflows and documentation provided by the authors.
Description
FAQ
Comments (8)
Thank you, it's a very organized and smooth workflow. I finally found a NSFW merge workflow that works flawlessly for me. It would be perfect if custom audio support were added to this workflow.
Definitely works nicely, once I juggled around a few things - smooth results overall! I am wondering where I'd be able to plug LTX-2 LoRAs in here though...I don't doubt that the phr00tmage LTX-2 merge includes a lot of the existing LTX-2 LoRAs in it, but I could see some extra strengthening being needed. Is this at all advisable, or is it better to just stick with the phr00tmage checkpoint as is?
Come to think of it, I also noticed this workflow doesn't appear to have any place for a negative prompt; the negative prompt is handled by a "ConditioningZeroOut" node it appears. Any particular reason for this workflow choice?
@Skunkylicious This was an oversimplified version of Lightricks’ distilled workflow. I will replace the prompt enhancer dependency and the multi‑stage LoRA process with their full workflow and post it here. The negative prompt is not required since we are using CFG 1 to speed up processing.
@FranciscoM Ah, fair enough, fair enough! Considering the results I've gotten visually I think the negative prompt wasn't strictly needed - just need to refine my positive prompts to get closer to what I'm after.
Phr00tmage merges are cool but not for scenarios just first frame action. If you prompt for a woman walking with a man to a bedroom they will thrust their way to the bedroom because of the furry or sex motion lora always active. If you want straight to the action then its cool but for scenarios its not very good. Someone needs to completely rewrite the LTX-2 model so sex motions activate when you want it to activate. Not to mention full blown nudity.
@jd666 So we're stuck waiting for an LTX-2 diffusion model that's trained from the ground up on a slew of NSFW videos of all stripes, not just base LTX-2 with LoRAs merged onto it huh? Hmmm. I've also noticed that the phr00tmage merges really don't like to handle rear views; sometimes when I've used it on a character seen from behind, their body will start to warp to look as if it's from a straight-on view, first the face, then the chest. It's...rather disturbing to watch.
@Skunkylicious I've added a workflow with LoRA support, but it's currently limited to 720p. Working with distilled versions is challenging because they are optimized for very few steps; the aggressive jumps between sigmas often introduce artifacts that interfere with both the LoRA detail and Lip-Sync.