Soft Enhance LoRa
This soft enhance LoRa for text to video, is a natural detail enhancer LoRA made for users who want a cleaner, more refined look. The text to video workflow can be found HERE just swap out the lora and prompt. All the sample vids are text to video.
It adds subtle detail, improves overall clarity, and helps soften some of the stronger contrast characteristics of LTX 2.3. The result is a smoother, more natural finish that keeps detail present without making the image feel overly sharp.
This is the softer counterpart to Crisp Enhance, which pushes a sharper, more cinematic beautifier-style look. Both LoRAs can be used together and balanced at different strengths to create a custom enhancer look depending on the scene and the amount of sharpness or softness you want.
This LoRa is Good for:
* Natural-looking enhancement
* Softer detail recovery
* Reducing harsh contrast
* Smoother skin and cleaner tonal transitions
* Dialing in a balanced look when mixed with Crisp Enhance
* Can be used on other styles, not just realism
Use Soft Enhance when you want added detail and polish, but with a gentler, more organic result.
Crisp Enhance LoRa
Crisp Enhance is a sharp detail enhancer LoRA made for users who want a cleaner, more defined, cinematic look.
It increases perceived sharpness, strengthens micro-detail, and enhances contrast to give images and video frames a more polished, high-definition appearance. The result is a crisp, cinematic finish that brings out texture and structure while adding a subtle beautifier-style clarity.
This is the sharper counterpart to Soft Enhance, which focuses on smoothing contrast and producing a more natural, organic look. Both LoRAs can be used together and balanced at different strengths to create a custom enhancer look depending on the scene and the level of sharpness or softness you want.
This LoRa is Good for:
Sharper detail enhancement
Cinematic-style clarity
Strengthening texture and edge definition
Beautifier-style polish
Dialing in a balanced look when mixed with Soft Enhance
Use Crisp Enhance when you want stronger detail, cleaner edges, and a more cinematic, high-definition result.
Description
Soft Enhance
Soft Enhance is a natural detail enhancer LoRA made for users who want a cleaner, more refined look.
It adds subtle detail, improves overall clarity, and helps soften some of the stronger contrast characteristics of LTX 2.3. The result is a smoother, more natural finish that keeps detail present without making the image feel overly sharp.
This is the softer counterpart to Crisp Enhance, which pushes a sharper, more cinematic beautifier-style look. Both LoRAs can be used together and balanced at different strengths to create a custom enhancer look depending on the scene and the amount of sharpness or softness you want.
This LoRa is Good for:
* Natural-looking enhancement
* Softer detail recovery
* Reducing harsh contrast
* Smoother skin and cleaner tonal transitions
* Dialing in a balanced look when mixed with Crisp Enhance
Use Soft Enhance when you want added detail and polish, but with a gentler, more organic result.
FAQ
Comments (80)
Damn good results from this one. Thank You!
Can it be used for i2v?And can it improve the face consistency?
When using i2v, u already start with an image so I don't think it will do much but u can try.
I still use the old LTX 2 "IC Detailer" Lora. Is this similar to that? I've been waiting for something that can replace it since I have a feeling it's maybe bad practice to use it with LTX 2.3.
I am not sure, I don't know what that was trained on. U can do comparisons and see what looks better
I think this might be the best LoRA ever made for LTX-2.3. Congratulations!
Hi, how are you? Last time I almost messaged you on Discord to ask if you had done anything interesting with LTX 2.3, and then right on cue I got notifications on CivitAI — I'm so glad you made these LoRAs, they're all amazing! But this one in particular I absolutely love. I left a sample video of a monster in the style of the movie The Thing — I always use that image for testing because it has no humanoid shape, which really helps reveal the model's limitations. And you've done an incredible job: it completely removes that clay/plasticine effect the distilled model left behind. I tested it and the videos are noticeably sharper now. Thank you so much for sharing these! 💝
This is really great for composition options, definitely a refinement in the shadow areas, the soft enhance lora is has a particularly nice general and professional feel. Vey good Please continue making these kind of tools. Easy Follow <3 ....
ps: just a strange heads up, i found some of your stuff on huggy a few days ago randomly, and i figured you'd be on here like most, so i tried a few key words linked to your loras, and also your name directly in the search up above here, and you didn't show up at all, nothing, so tried a few times, nope, it insisted there was no such person.... so i gave up... then just now today, out of the blue - booP - there you are, and i see you've been here for years ... well, could have been a glitch... but your name was not directly searchable for me until just now .... strange eh
thank you for you work
Glad u like it and yes very strange about not being able to find me
What workflow are you using vrgamedevgirl? because I feel my videos do not come out as clear as your examples
Pm me on Discord vrgamedevgirl
Great work!! Was the training solely based on images, or were videos also used? And how many images or videos were used in the training? Thanks!
If there is a reply, please let me know. I'm also very curious about whether it was trained based on a high-quality image dataset or a video dataset. Thank you.
Yes images only. About 50
@vrgamedevgirl Thank you! I really like this loras!
Definitely amazing... if someone starts working on fixing the atrocious audio on LTX2.3, then we'll be golden.
You were heard by the gods — they just released version 1.1 of the distilled model, which improves the audio.
https://huggingface.co/Lightricks/LTX-2.3
@lug_L Nice! Once a GGUF is created i'll be trying this. For some reason GGUF always just looks better for me.
@vrgamedevgirl Kijai already has the 1.1 model in case you want to try it. For some reason his models work faster for me, the GGUF ones run slower.
https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/diffusion_models
@lug_L Thanks, GGUF might be a bit slower but it always looks better for me. But maybe i'll try this one thanks!
@lug_L Holy... I'm really starting to believe solipsism is real.. thanks for the heads up!
😍 👍🏻
Excellent work. Could you please provide the I2V workflow? Thank you.
The text to video workflow is in the description. This is not meant for image to video since you already start out with an image when using I2V this really won't have an effect. All the sample vids are text to video.
what strengh for the crisp lora? 1?
I have used 1, both with t2v and i2v. I think it works very nicely!
1.0
@Just_Claire when using for image to video is it actually doing anything?
@vrgamedevgirl Yes definitely! This morning I had one vid running without your wonderful Lora and it was not as sharp and crisp as running it with it!
@Just_Claire with other loras too at 1?
@dominik1996d926 I also use the 'LTX 2.3 - Fantasy Realism' Lora for i2v at about 0.65 to enhance the (female) characters (they get prettier and also get more beautiful hair) and with animals their details and fur, scales, etc. get a lot better. With the 'LTX 2.3 - Luxe Sensual' Lora (also with i2v at about 0.65) the details get richer and the female character gets that 'glamour vibe', and a more consistent character resembling the character you can see in the promo videos of the creator vrgamedevgirl. (I just LOVE those three Loras!)
@Just_Claire and what about fantasy realism and crisp lora on 1?
@dominik1996d926 The 'Crisp' lora enhances sharpness and details both with t2v and also itv (I have it always on 1.0) I will try the Fantasy Realism lora on i2v on 1.0 and will let you know :)
(EDIT: ..a few hours later.. The Fantasy Realism lora (set at 1.0) on i2v seems to enhance quality of fingers and hands a bit when they are moving, also more detail in clothing and a slight change of character (really not that much). )
@Just_Claire can i combine the crisp lora with lora detailer lora?
@dominik1996d926 I have combined the Crisp lora with the Luxe Sensual lora (0.5 - 1.0) and also with the Fantasy Realism lora (0.5 - 1.0) successfully, so yes, you can. :)
You are the Queen of Civitai! I'm not even exaggerating! this is Gem! thank you so much!
Thank you for this Lora, but the download link leads to an error."error getting file"
That sounds like a civitai bug.
@vrgamedevgirl I managed to retrieve it from Hugging Face. Great work, thanks again ;)
@ricoss35852 Which one from hugging face though? I had two there and one of them is not published here and the names are different. I plan to update hugging face soon though.
@vrgamedevgirl I went to the Civarchive website and found the link hugging face.
"https://huggingface.co/Sentinel7/ltxv/resolve/main/2535622/2849706/LTX2.3_Soft_Enhance.safetensors"
Is this racial agnostic or more biased towards caucasian features ? In the example vid the lora was making the face very caucasian coded or straight up changing the race.
it was trained on a variety of faces and different ethnicities. I'll take a look.
what example video are you referring to?
@vrgamedevgirl at 0:21 and 0:38 for both the black girls.
P.S. I dunno if over the internet the original comment came across with additional undertones. The intended question was asked from purely a technical info perspective to make an informed choice in workflows.
@benshapiro679 no worries I understood what you meant. :) I'm still not sure what videos though. The sample videos I created, the I'm not seeing what your saying. They follow the prompt unless I'm missing something.
@vrgamedevgirl i was talking about the youtube video that is in crisp enhance page. at the timestamps 0:21 and 0:38 the lora was darker complexion and soft.lora was more caucasian.
@benshapiro679 ahhh the prompt didn't have an ethnicity in it so LTX kind of just picked one I think, not sure.
Thank you for this lora, I wonder if there is a boost for I2V
I have not tested that but others claim there is. Give it a shot!
Hi Gamedevgirl, great to see you over on the LTX stuff. your work with fusionX gave me so much joyful results in WAN.
I'm using wan2gp UI to run LTX2.3 so i can't import your workflow. Could you let me know if you created these LORAs to be run at 1.0 on the initial phase and also 1.0 on the second phase (upscaler phase).
For example the old IC detailer used to be usually applied to just the second phase to improve fine details.
There is also people using just the initial phase but all the way up to final resolution.
I'm not sure... any info would be greatly appreciated as i don't use comfyui so can't check your workflow.
In all my sample vids I use 1.0 on both stages :)
It really works great! Thank you.
The quality is insane.
ABSOLUTE CINEMA
Why everyone's ltx videos look so nice and smooth and crisp, and mine look blurry with extra hands and things morphing/disappering? IT's so frustrating, I already tried multiple workflows, tried Q5, Q8, same thing, I can hardly put one barely acceptable D:
yeah idk im asking the same question
try cfg 1.1 or higher, try a different sampler, euler_a or res_multistep, you could also try putting the new lora right after the checkpoint and before any other lora, turn the lora strength down to .5 and test from there, basically if you are getting weird or chaotic images/vids its a configuration issue or a few settings that need adjustment,
I get best results with it first before the Omni lora
I use a distilled GGUF ltx model and have found that is the only way to get good results.
@vrgamedevgirl Or using Kijai's distilled LoRA's with dev models works as well. Pre-distilled GGUF's leave something to be desired when it comes to fluid motion.
@betech79209 actually it has been tested and the Q8 GGUF does just as good as the main model and better than FP8.
@vrgamedevgirl I'll reserve judgement on that. Some of the GGUF's I have tested do restrict movement or remove certain details that were in the original BF16. Some also leave 'creator' watermarks from those quantized the model.
keep getting noise/ static outputs
on first pass drop the strength to 0.50 and 2nd pass do 1.00
If you are not using a distilled LoRA before the enhancement LoRA, you will get noisy audio. Something like Kijai's Dynamic fro09_111 distilled LoRA should be the first LoRA in the chain. Use Kijai's LoRA with a strength of 0.60.
That being said, you never stated if you were using a dev model or a distilled model.
please use the workflow i provided.
@vrgamedevgirl Provided workflows only work for a certain hardware combinations. They are not one-size fits all. I've had to tweak other peoples' workflows because they were either scam-flows or didn't output the quality that was expected.
@betech79209 My Workflow is very basic and should work for anyone. NO crazy custom nodes or anything. And not a scam... I would never do that. I'm well known in the AI and comfyui community and would never do that.
can i use your crisp lora alongside lora detailer?
yes you can mix them but i would not do both at 1.0
I use it on ltx2.3 1.1
keep getting this from the first workflow RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 4096 but got size 3840 for tensor number 1 in the list.
The model that you are using is not compatible with the LoRA.
is this supposed to be for T2V only? Because i cant spot any difference when using it with I2V tbh
yes its mainly for text to video as image to video you have a starting image already, its not an IC lora
does this only work with close up faces?
No, it can work on anything
I see no difference with this other than it tries to add nudity when not wanted.
must be your prompt as I and others have not had this issue. Are you using the workflow in the link? can you share your prompt? The dataset had NO nudity at all. Not even partial.
What main model are you using?
Details
Files
LTX2.3_Soft_Enhance.safetensors
Mirrors
LTX2.3_Soft_Enhance.safetensors
LTX2.3_Soft_Enhance.safetensors
LTX2.3_Soft_Enhance.safetensors
LTX2.3_Soft_Enhance.safetensors
LTX2.3_Soft_Enhance.safetensors
LTX2.3_Soft_Enhance.safetensors
LTX2.3_Soft_Enhance.safetensors
LTX2.3_Soft_Enhance.safetensors
sftnhnc.safetensors
LTX23_Soft_Enhance.safetensors
LTX2.3_Soft_Enhance.safetensors
LTX2.3_Soft_Enhance.safetensors
LTX2.3_Soft_Enhance.safetensors
Soft Enhance - LTX2.3.safetensors
LTX2.3_Soft_Enhance.safetensors
LTX2.3_Soft_Enhance-mid_2535622-vid_2849706.safetensors
LTX2.3_Soft_Enhance.safetensors
Soft Enhance - LTX2.3.safetensors
Soft Enhance - LTX2.3.safetensors
LTX2.3_Soft_Enhance.safetensors
LTX2.3_Soft_Enhance-mid_2535622-vid_2849706.safetensors
LTX2.3_Soft_Enhance.safetensors
LTX2.3_Soft_Enhance.safetensors
LTX2.3_Soft_Enhance.safetensors
Enhancers Soft Enhance - LTX2.3 - adds detail and polish, but with a gentler, more organic result.safetensors
Soft Enhance - LTX2.3.safetensors
Soft Enhance - LTX2.3.safetensors
Enhancers Soft Enhance - LTX2.3 - adds detail and polish, but with a gentler, more organic result.safetensors
LTX2.3_Soft_Enhance.safetensors