TL;DR
CinEro Illustrious V5 is a further development of my Pony / Illustrious models designed for better realism and versatility. Highlights:
- High resolution: tested with 1st pass rendering at 1440x768 resolution without high chance to get duplicate elements.
- Enhanced realism: Improved color accuracy, contrast, and sharpness for lifelike images.
- Vintage cinematic aesthetic: Analog film grain, vintage lens blur, and natural imperfections.
- Two versions: FP8 (optimized for RTX GPUs with <12GB VRAM) and FP16 (stable, high-quality CLIP/UNet).
- FP8 performance: Faster inference with minimal resource usage (requires ComfyUI FP8 mode).
- All version have VAE embedded. This VAE is just a standard SDXL VAE with some memory optimizations.
Support the CinEro Project
Training CinEro Illustrious V5 cost me over 10,300 BUZZ, and your support can help fuel future improvements! Here’s how you can contribute:
1. Donate BUZZ: Every contribution goes directly into training the next iteration of CinEro.
2. Spread the word: Share the model on social media, recommend it to friends, and engage with the community.
3. Post your creations: Upload images generated with CinEro, interact with others, and help grow the model’s popularity.
Your time and support are invaluable—whether through BUZZ donations or community engagement, you’re helping shape the future of CinEro!
Recommended Parameters & Tips
[TBD]
Acknowledgments
A huge thank you to everyone who supported previous versions of CinEro through donations, feedback, comments, and sharing their amazing creations. Your contributions have been instrumental in bringing CinEro Illustrious V5 to life.
List of kind supporters:
~={ @belember }=~ ~={@SerLT}=~ ~={@KainKarnage}=~ ~={@Maximo666}=~
@HcupMania @AFD_0 @matrix2014 @AiBan @gman_umscht @HK700 @mphobbit @warlords @namenoname @antigen3 @MysticMindAi @acaciahoney @iheartlumina @jakobhovman259 @funkyflame42513 @Brewce @TheSmokingGPU
@Lizardon1025
Description
FP16 (stable, high-quality CLIP/UNet)
FAQ
Comments (22)
I'm currently using a LoRA created from your released checkpoint model. However, after trying the new version, I've noticed that the generated images aren't as clear and defined as the previous version—similar to what I described here: https://civitai.com/posts/14158894. To double-check, I generated images without applying my LoRA, and the results were like this: https://civitai.com/posts/14158806.
Compared to the previous version, the faces now appear more Westernized and somewhat blurred overall. Is this the intended direction of the checkpoint model going forward? Given that the style has shifted significantly from v4, I don't think I'll be able to update my characters or images to this new version.
Thanks to hard job, and expecting future good results.
"Westernized" - yes. My dataset mostly contains EUR characters. But some of the characters captioned as "asian" by the WD1.4 tagger. I suspect if the character looks like Mid-East person and marked as "asian" it can damage a real asian segment of the model.
Hopefully I will prepare the new datasets with more balanced nationality. Sorry that this release disappointed you.
As a quick fix I can suggest you to mix in CLIP part from some model like Hassaku with strong anime bias with better representation of asian nationalities.
Love the CinEro models, this one is a step above. Great work!
smth to improve?
face expressions, poses?
@homoludens I will have to play around with it some more, currently training some LoRAs on it!
I love the outputs of this model with the only exception that all my images require brightness, contrast, shadows rebalancing now matter what settings I change in the workflow.
To which direction adjust brightness / contrast?
I always have to increase brightness, contrast, shadows, sometimes saturation but not always. The images come out dim and under saturated. It could be just a matter of preference, no real reason to make changes just noting the observation.
@Visidious99 Saturation and contrast is tricky... In reality common images taken with mobile camera or taken by low- or mid- grade camera / lenses have lower contrast and saturation. Modern Computational Photography in smartphones replaces the colors and gain contrast / saturation which make image digital and unreal.
I am thinking hard how to combine the colors / contrast from CinEro V3 / V4 and for example aesthetics from SatPony series with sharpness, realism and realistic skin textures. It is complex problem. But I'm still trying.
Thanks for feedback. It helps to progress.
@Visidious99 Brightness was lowered - I confirm what you stated.
More adjustments needed.
Hey there, 2 questions:
Why this is not with the other version 4?,like why separated
And how is this different from the other Cinero V3 sin?
1. This is separate for few reasons. I decided to change permissions (restriction on merges looks nonesense as impossible to detect or prevent). Also, I feel that there can be more than one FP16 version (fixes, etc.) and I planned also release the FP8 versions for Low Spec computers. Stacking all of them in one model seems inconvenient.
2. V2 is a radically realistic but with major flaw - HDR-like rendering. V3 was the 1st attempt to combine realism and sharpness from V2 and general cinematic style of the whole CinEro family (unsuccessful or partial success). V4 much closer to what I want to achieve, but still have the flaws (lower realism, more like Semi-Real Anime models). V5 is a swing back toward V2 level of realism with cinematic flavor, but again have it's flaw - on a single pass renders it gives flawy eyes . 2nd pass like with HiRes Fix corrects most of the issues, but not all users prefer quality over the speed. You can use Step 24 for the first pass + Step 8 for HiRes Fix, but somebody can feel it too much. So, I need more time and BUZZ to improve V5 further.
@homoludens damn, thank you so much for the full info🙌
@homoludens I have noticed that eye problem on so many checkpoints. Either they're too small and a squashed mess, or too vacant to have any expression. Animagine is the only checkpoint I've seen that consistently gives amazing character and expression to faces without prompting. Adding a tag like "assertive_female" and you can see the effect straight away, it never gives blank stares. But how to get this with cyber realistic XL quality eyes in one checkpoint?
@Tootle Thanks for mentioning it... I can try to merge only the textual part of the Animagine (CLIP module) and check if mentioned tag works.
10k buzz? Or you are just joking?
Two LORAs 4.3k BUZZ on CivitAI + bunch of draft images and few days of local GPU work,
You mention fp8, but I'm not seeing it in this version. Do you plan to release it?
Are you interested in it? It is easy to make it, but eventually I didn't notice very big interest in it.
I can release it soon, but have no time to render the samples. It would be good if someone test it and post some images.
Hey! Uploading FP8. Should be completed in 1-2 hrs.
@homoludens Thank you! I'll give it a go when I see it up.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.